Link¨ oping Studies in Science and Technology

Dissertation No. 1264
Iterative Filtered Backprojection
Methods for Helical Cone-Beam CT
Johan Sunneg˚ardh
Department of Electrical Engineering
Link¨ opings universitet, SE-581 83 Link¨ oping, Sweden
Link¨ oping August 2009
Link¨oping Studies in Science and Technology
Dissertation No. 1264
Author
Johan Sunneg˚ardh
Department of Electrical Engineering
Link¨ oping University
SE-581 83 Link¨ oping, Sweden
Copyright c ( 2009 Johan Sunneg˚ardh
Sunneg˚ardh, Johan
Iterative Filtered Backprojection Methods for Helical Cone-Beam CT
ISBN 978-91-7393-586-9
ISSN 0345-7524
Typeset with L
A
T
E
X
Printed in Sweden by LiU-tryck, Link¨ oping, 2009
iii
Abstract
State-of-the-art reconstruction algorithms for medical helical cone-beam Com-
puted Tomography (CT) are of type non-exact Filtered Backprojection (FBP).
They are attractive because of their simplicity and low computational cost, but
they produce sub-optimal images with respect to artifacts, resolution, and noise.
This thesis deals with possibilities to improve the image quality by means of iter-
ative techniques.
The first algorithm, Regularized Iterative Weighted Filtered Backprojection
(RIWFBP), is an iterative algorithm employing the non-exact Weighted Filtered
Backprojection (WFBP) algorithm [Stierstorfer et al., Phys. Med. Biol. 49, 2209-
2218, 2004] in the update step. We have measured and compared artifact reduction
as well as resolution and noise properties for RIWFBP and WFBP. The results
show that artifacts originating in the non-exactness of the WFBP algorithm are
suppressed within five iterations without notable degradation in terms of resolution
versus noise. Our experiments also indicate that the number of required iterations
can be reduced by employing a technique known as ordered subsets.
A small modification of RIWFBP leads to a new algorithm, the Weighted
Least Squares Iterative Filtered Backprojection (WLS-IFBP). This algorithm has
a slightly lower rate of convergence than RIWFBP, but in return it has the attrac-
tive property of converging to a solution of a certain least squares minimization
problem. Hereby, theory and algorithms from optimization theory become appli-
cable.
Besides linear regularization, we have examined edge-preserving non-linear reg-
ularization. In this case, resolution becomes contrast dependent, a fact that can
be utilized for improving high contrast resolution without degrading the signal-to-
noise ratio in low contrast regions. Resolution measurements at different contrast
levels and anthropomorphic phantom studies confirm this property. Furthermore,
an even more pronounced suppression of artifacts is observed.
Iterative reconstruction opens for more realistic modeling of the input data
acquisition process than what is possible with FBP. We have examined the pos-
sibility to improve the forward projection model by (i) multiple ray models, and
(ii) calculating strip integrals instead of line integrals. In both cases, for linear
regularization, the experiments indicate a trade off: the resolution is improved at
the price of increased noise levels. With non-linear regularization on the other
hand, the degraded signal-to-noise ratio in low contrast regions can be avoided.
Huge input data sizes make experiments on real medical CT data very demand-
ing. To alleviate this problem, we have implemented the most time consuming
parts of the algorithms on a Graphics Processing Unit (GPU). These implementa-
tions are described in some detail, and some specific problems regarding parallelism
and memory access are discussed.
iv
v
Popul¨arvetenskaplig sammanfattning
Datortomografi (f¨ orkortat DT eller CT), ¨ aven kallad skiktr¨ ontgen, ¨ ar en vanligt
f¨orekommande teknik inom medicinsk bilddiagnostik f¨ or att avbilda kroppens inre.
Till skillnad fr˚an en vanlig r¨ ontgenunders¨ okning som anv¨ ander r¨ ontgenstr˚alar i en
enda riktning, anv¨ ander CT r¨ ontgenstr˚alar i flera olika riktningar f¨ or att rekon-
struera bilder av skikt i kroppen. En s˚a kallad rekonstruktionsmetod erfodras f¨ or
att kunna ber¨ akna bilderna fr˚an uppm¨ atta data.
Alla rekonstruktionsmetoder som idag anv¨ ands f¨ or medicinsk datortomografi
¨ ar icke-exakta. Detta betyder att de framr¨ aknade bilderna inneh˚aller fel oavsett
noggrannhet och uppl¨ osning vid datainsamlingen. F¨ or ¨aldre tomografer med f˚a
detektorrader ¨ ar dessa fel f¨ orsumbara, men p˚a nyare mer effektiva tomografer med
64 eller fler detektorrader kan st¨ orande artefakter uppst˚a. I denna avhandling un-
ders¨oks olika iterativa metoder f¨ or att undertrycka dessa artefakter, samt f¨ orb¨ attra
de rekonstruerade bilderna med avseende p˚a sk¨arpa och brus.
Den f¨orsta metoden, som p˚a engelska f¨orkortas RIWFBP, bygger p˚a iterativ
till¨ ampning av den icke-exakta metoden WFBP [Stierstorfer et al., Phys. Med.
Biol. 49, 2209-2218, 2004]. F¨ or RIWFBP har reduktion av artefakter, brus-
niv˚aer samt spatiell uppl¨ osning m¨atts genom rekonstruktion av f¨ or ¨andam˚alet
konstruerade testobjekt. Resultaten visar en tydlig minskning av artefaktniv˚an
under de f¨ orsta fyra iterationerna. En negativ bieffekt ¨ ar att bildernas brusniv˚a
okontrollerat ¨ okar med antalet iterationer. F¨ or att f¨ orb¨ attra konvergensegenskaper
och undvika denna bieffekt f¨ oresl˚as och unders¨oks s˚a kallad linj¨ ar regularisering,
vilket inneb¨ ar att de rekonstruerade bilderna i varje iteration filtreras med ett
linj¨ art l˚agpassfilter.
Ist¨allet f¨or att, som i RIWFBP, anv¨ anda alla tillg¨ angliga m¨ atningar i varje
iteration, kan en delm¨ angd anv¨ andas. Detta utg¨ or grunden till metod nummer
tv˚a, f¨ orkortad OS-IWFBP. Det visar sig att antalet n¨ odv¨ andiga iterationer sjunker
n˚agot n¨ar denna teknik anv¨ ands. En nackdel ¨ ar dock att brusniv˚an ¨okar j¨ amf¨ort
med RIWFBP.
En nackdel med RIWFBP ¨ ar att den till skillnad fr˚an de flesta andra iterativa
rekonstruktionsmetoder inte enkelt kan beskrivas som l¨ osningen till ett minimer-
ingsproblem. I avhandlingen f¨ oresl˚as en tredje metod, f¨orkortad WLS-IFBP, som
med en gradientbaserad metod minimerar en kvadratisk m˚alfunktion. Den nya
metoden undertrycker artefakter ungef¨ ar lika effektivt som RIWFBP-metoden,
men har f¨ ordelen att den kan analyseras och utvidgas med v¨ alk¨anda verktyg fr˚an
omr˚adet optimeringsl¨ ara.
Med avseende p˚a sk¨arpa kontra brus ¨ ar den linj¨ art regulariserade RIWFBP-
metoden likv¨ ardig med WFBP. F¨ or att f¨ orb¨ attra f¨ orh˚allandet mellan sk¨ arpa och
brus har ¨ aven icke-linj¨ ar regularisering unders¨ okts, vilket betyder att de rekon-
stuerade bilderna i varje iteration filtreras med ett icke-linj¨ art kantbevarande filter.
Experiment visar att denna typ av regularisering kan f¨ orb¨ attra den sk¨ arpan f¨ or
h¨ogkontrastdetaljer (ben) utan f¨ ors¨amring av signal/brus-f¨ orh˚allandet i omr˚aden
vi
med l˚aga kontraster (mjukdelar).
Till skillnad fr˚an direkta metoder m¨ ojligg¨ or iterativa metoder s˚asom RIWFBP
och WLS-IFBP modellering av r¨ ontgenk¨ allan och detektorelementens storlekar,
samt systemets r¨orelse under datainsamlingen. Vi har j¨ amf¨ort rekonstruerade
bilder med och utan noggrann modellering av dessa egenskaper. I det linj¨ art
regulariserade fallet visar resultaten att noggrann modellering ger b¨ attre sk¨arpa
till priset av ¨ okade brusniv˚aer. Med hj¨ alp av icke-linj¨ ar regularisering kan denna
kompromiss delvis undvikas.
P˚a grund av stora datam¨ angder ¨ ar experiment p˚a riktiga medicinska pro-
jektionsdata mycket kr¨ avande. F¨ or att minska ber¨ akningstiderna har de mest
ber¨akningskr¨ avande operationerna implementerats p˚a grafikh˚ardvara. I avhan-
dlingen beskrivs dessa implementationer samt ungef¨ arliga prestandavinster. Vi-
dare diskuteras specifika problem r¨ orande parallelisering och minnes˚atkomst.
vii
Acknowledgments
I would like to take this opportunity to thank all those who have contributed to
this thesis.
First and foremost, Per-Erik Danielsson who introduced me to the field, for his
never ending commitment and enthusiasm, for all advice concerning the writing,
and for always leaving his door open.
Michael Felsberg for all his support and advice in the final stages of my work, for
proof-reading, and for many interesting technical discussions.
G¨osta Granlund for welcoming me to work and study at the Computer Vision
Laboratory.
Maria Magnusson for interesting discussions, for proof-reading, and for providing
answers to my never ending flow of questions regarding various topics in the field
of image processing and computed tomography.
Fredrik Larsson, Johan Hedborg, and Klas Nordberg for proof-reading and pro-
viding valuable questions and comments.
All past and present members at the Computer Vision Laboratory for creating a
friendly and inspiring atmosphere.
Karl Stierstorfer, Siemens Healthcare, for many valuable comments and questions
regarding my work, for providing simulation and reconstruction software, and for
being an invaluable source of information about CT.
Herbert Bruder, Siemens Healthcare, for many interesting discussions, and for
proof-reading the manuscript.
Thomas Flohr, Siemens Healthcare, for inviting me to work and study at Siemens
in Forchheim, Germany.
All other people that I learned to know at Siemens during my stay in Germany,
for making me feel welcome, and for many interesting discussions.
Tommy Elfving and Lars Eld´en for sharing their knowledge in numerical analysis
and optimization theory.
Michel Defrise for many useful questions and comments on my licentiate thesis.
Maria Antonson and Petter Quick, CMIV, for providing excellent photos of a
modern CT system.
Ulrike for all the love and support you have given me, and for making me happy.
My family for all support during the writing of this thesis.
The financial support from Siemens Healthcare is gratefully acknowledged.
viii
Contents
1 Introduction 1
1.1 Computed Tomography (CT) . . . . . . . . . . . . . . . . . . . . . 1
1.2 Reconstruction methods and artifacts . . . . . . . . . . . . . . . . 2
1.3 Iterative Filtered Backprojection (IFBP) . . . . . . . . . . . . . . . 4
1.4 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Contributions and publications . . . . . . . . . . . . . . . . . . . . 7
1.6 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Analytical reconstruction methods 9
2.1 Two-dimensional image reconstruction . . . . . . . . . . . . . . . . 9
2.2 Weighted filtered backprojection . . . . . . . . . . . . . . . . . . . 13
3 Iterative reconstruction methods 19
3.1 A basic example. Application overview. . . . . . . . . . . . . . . . 19
3.2 Different types of iterative methods . . . . . . . . . . . . . . . . . . 21
3.3 The projection operator . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.1 Matrix representation . . . . . . . . . . . . . . . . . . . . . 23
3.3.2 Basis functions . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.3 Irradiation functions . . . . . . . . . . . . . . . . . . . . . . 26
3.3.4 The Joseph forward projection method . . . . . . . . . . . . 30
3.4 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4 Simulation and evaluation 35
4.1 Generation of input projection data . . . . . . . . . . . . . . . . . 35
4.2 Phantoms and error/noise measures . . . . . . . . . . . . . . . . . 37
4.3 Spatial resolution measurements . . . . . . . . . . . . . . . . . . . 39
5 Regularized iterative weighted filtered backprojection 45
5.1 Description of the reconstruction algorithm . . . . . . . . . . . . . 45
5.2 The regularization operator R . . . . . . . . . . . . . . . . . . . . . 48
5.3 Local region of interest (LROI) . . . . . . . . . . . . . . . . . . . . 49
5.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.4.1 Artifacts and noise measurements . . . . . . . . . . . . . . 50
5.4.2 Spatial resolution . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4.3 Higher number of iterations. Long objects. . . . . . . . . . 60
5.4.4 Local region of interest . . . . . . . . . . . . . . . . . . . . . 63
5.5 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . 65
x Contents
6 Ordered Subsets IWFBP (OS-IWFBP) 71
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2 The Ordered Subsets (OS) technique . . . . . . . . . . . . . . . . . 71
6.3 OS-IWFBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.5 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . 79
7 Relation between RIWFBP and least squares methods 83
7.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.2 Weighted least squares IFBP (WLS-IFBP) . . . . . . . . . . . . . 84
7.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.3.1 Artifact reduction . . . . . . . . . . . . . . . . . . . . . . . 87
7.3.2 Spatial resolution and noise . . . . . . . . . . . . . . . . . . 88
7.4 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . 90
8 Nonlinear regularization 95
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.2 Description of the reconstruction method . . . . . . . . . . . . . . 95
8.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
8.3.1 Spatial resolution versus local contrast . . . . . . . . . . . . 98
8.3.2 Edge versus surface based resolution measurements . . . . . 101
8.3.3 Artifacts and noise . . . . . . . . . . . . . . . . . . . . . . . 104
8.4 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . 108
9 Data acquisition modeling 113
9.1 Irradiation functions . . . . . . . . . . . . . . . . . . . . . . . . . . 113
9.2 Multiple ray forward projection . . . . . . . . . . . . . . . . . . . . 114
9.2.1 An improved acquisition model . . . . . . . . . . . . . . . . 114
9.2.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 117
9.2.3 Discussions and conclusions . . . . . . . . . . . . . . . . . . 123
9.3 A strip integral projection method . . . . . . . . . . . . . . . . . . 127
9.3.1 Method description . . . . . . . . . . . . . . . . . . . . . . . 127
9.3.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 130
9.3.3 Discussions and conclusions . . . . . . . . . . . . . . . . . . 131
10 Implementation on Graphics Processing Units (GPU) 135
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
10.2 The GPU from a CUDA perspective . . . . . . . . . . . . . . . . . 136
10.3 Programming CUDA . . . . . . . . . . . . . . . . . . . . . . . . . . 138
10.3.1 Kernels and threads . . . . . . . . . . . . . . . . . . . . . . 138
10.3.2 Thread organization . . . . . . . . . . . . . . . . . . . . . . 139
10.3.3 Textures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
10.3.4 Constant memory . . . . . . . . . . . . . . . . . . . . . . . 142
10.4 Forward projection . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.5 Backprojection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
10.5.1 The WFBP/standard backprojector . . . . . . . . . . . . . 145
10.5.2 The Joseph backprojector P
T
. . . . . . . . . . . . . . . . . 147
10.6 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . 149
Contents xi
11 Conclusions and future research 151
11.1 Summary of conclusions . . . . . . . . . . . . . . . . . . . . . . . . 151
11.2 Suggestions for future research . . . . . . . . . . . . . . . . . . . . 153
Appendices 155
A Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
B Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
References 158
xii Contents
Chapter 1
Introduction
1.1 Computed Tomography (CT)
In the areas of medical diagnostics and non-destructive testing, it is of great inter-
est to be able to reproduce images of the interior of objects. One common technique
to accomplish this feat is known as Computed Tomography (CT), which originates
from 1972, when Sir Godfrey Hounsfield at Electrical and Musical Industries lim-
ited (EMI) patented the first CT scanner [118]. The technology became a huge
success and in 1979, Alan Cormack and Hounsfield were awarded the Nobel Prize
for this invention.
A CT scanner uses digitally sampled X-ray images acquired in multiple di-
rections to calculate cross-sectional images of the X-ray attenuation of an object.
High acquisition speed is generally desirable since this reduces examination time
and image artifacts due to motion. Since 1972, several major improvements have
been made in this direction. A comprehensive survey of early developments in CT
is given by Webb [115]. Later developments are covered in the books by Kalender
[46] and Hsieh [37]. Here, we give a brief presentation of the state of the art at
the time of writing.
Fig. 1.1 shows a photograph of a modern CT scanner. The data acquisition
system, consisting of an X-ray source and a detector, is attached to the gantry
hidden inside the main part of the scanner. Thanks to the so called slip ring
technique for transfer of electrical power and data, no cables need to be attached
to the gantry. Previously, the gantry was supported by cables which typically
tolerated one turn after which the gantry had to change its direction of rotation.
During an examination, the slip ring connected gantry is rotating with a con-
stant speed of between one and four turns per second. Different regions of the
patient are irradiated by translating the patient table along its z-axis, i.e. orthog-
onal to the gantry rotation plane. Depending on this motion, different scanning
trajectories are obtained. The most common one is the helical trajectory, obtained
by translating the patient table with constant speed during data acquisition.
2 Introduction
z
Figure 1.1: A modern medical CT scanner (by courtesy of Center for Medical
Image Science and Visualization (CMIV), Link¨ oping, Sweden).
Fig. 1.2 illustrates the helical scanning geometry in the coordinate system of
the examined object. The detector is shaped as a cylinder centered around the
source and aligned with the z-axis. Its height h when projected onto the center of
rotation, or equivalently, the cone-angle κ
max
, determines an upper limit on the
helical pitch P approximately equal to
P
max
≈ 1.4h = 1.4 2R
S
tan κ
max
, (1.1)
where R
S
is the distance from the source to the center of rotation.
During the past decade, many developments towards higher acquisition speeds
have been introduced by increasing the number of detector rows and improving
systems for transferring data from the gantry. In 1998, the first scanners with four
detector rows were released, and three years later, in 2001, the number of detector
rows in high end scanners was increased to sixteen [46]. Today, 64 row detectors
are quite common. In combination with other techniques for accelerating data
acquisition, such as dual source systems, it is now possible to perform a full body
examination in less than five seconds.
1.2 Reconstruction methods and artifacts
An important part of any CT system is the reconstruction method, which trans-
forms the measured data into images. Reconstruction methods for CT can be ei-
ther analytical or iterative. An analytical method is based on exact or approximate
mathematical inversion of the projector P
phys
that models the data acquisition pro-
cess. In contrast, iterative methods do not involve explicit inversion, but use an
iterative procedure to generate a sequence of reconstruction results converging to
1.2 Reconstruction methods and artifacts 3
X-ray source
Detector
Projection of the detector onto the
center of rotation (isocenter)
x
y
z
2
κm
a
x
P
h
Figure 1.2: Scanning geometry for helical cone-beam CT. To simplify the visu-
alization, the detector height h and the pitch P of the helix are exaggerated.
an image or volume f that in some sense minimizes the difference between the pro-
jection data P
phys
f and the measured data. These things will be explained more
precisely in Chapter 2 and 3. Although the EMI scanner by Hounsfield employed
an iterative method [118], reconstruction methods in commercially available CT
systems are mainly analytical. One reason for this is that analytical methods are
computationally much less expensive than iterative methods.
Analytical methods can be either exact and non-exact. Exact methods are
based on exact inversion of the projector P
phys
in the continuous domain. Thus,
in an ideal scanning situation (monochromatic radiation, linear detector response,
and no geometric inaccuracies), reconstruction results can be made arbitrarily ac-
curate by using sufficiently high sampling densities and tube currents. Although
efficient exact methods based on work by Katsevich [48] and Zou and Pan [128] ex-
ist, they are currently not found in clinical use. Instead, manufacturers of clinical
CT systems employ non-exact methods, or more specifically, Filtered Backprojec-
tion (FBP) methods [58, 44, 111, 26, 100]. Due to approximations in the derivation
of these methods, reconstruction results are contaminated by cone artifacts, the
name stemming from the fact that these artifacts become more pronounced for
higher cone angles. In return, non-exact methods are computationally less de-
manding, simpler to implement, and offer a better dose utilization than exact
methods.
One state-of-the-art analytical non-exact reconstruction method in helical
4 Introduction
x
y
Figure 1.3: An axial image of a clock phantom reconstruction distorted by cone
and windmill artifacts. The cone artifacts are slowly varying, and appear as dark
and bright regions in the vicinity of the spheres. The windmill artifacts appear as
multiple dark and bright streaks tangent to the spheres. Greyscale window ±20
Hounsfield Units (HU).
cone-beam CT is the Weighted Filtered Backprojection (WFBP) by Stierstor-
fer et al. [100]. Fig. 1.3 shows an axial image of a WFBP reconstruction of the
sphere clock phantom by Turbell [110], specially designed for examining the non-
exactness of a reconstruction method. The image shows two types of artifacts:
windmill artifacts and cone artifacts. Wind-mill artifacts are caused by aliasing in
the z-direction of the detector, and look like alternating dark and bright streaks ex-
tending radially from the upper and lower ends of the spheres. The cone-artifacts
are more slowly varying than the windmill artifacts. In Fig. 1.3, they show up as
one or two pairs of dark and bright regions extending from the spheres. Another
example of cone artifacts and windmill artifacts is given in Fig. 1.4, showing a
sagittal image of a WFBP reconstruction of a physical head phantom.
1.3 Iterative Filtered Backprojection (IFBP)
The purpose of the work presented in this thesis is to investigate the possibilities
to improve FBP methods with respect to artifact reduction, spatial resolution,
and signal to noise ratio, by means of Iterative Filtered Backprojection (IFBP).
Below, we give a brief presentation of a simple IFBP method. A more detailed
presentation and analysis is given in Chapter 5 and 7.
1.4 Thesis outline 5
z
y
Figure 1.4: A sagittal image of a physical head phantom reconstruction distorted
by cone and windmill artifacts. The dashed circle and ellipse indicate windmill
artifacts, and the solid ellipse indicates a cone artifact.
Fig. 1.5 illustrates the basic principle of IFBP. Suppose that the input data
represented by the vector p
in
∈ R
M
have been obtained from the object of interest
f ∈ L
2
(R
3
) by the linear operator P
phys
: L
2
(R
3
) → R
M
modeling the data ac-
quisition process. Furthermore, let the matrix Q ∈ R
N×M
represent a non-exact
FBP method (N is the number of reconstructed voxels). An approximation of the
correct result is then given by f
0
= Qp
in
∈ R
N
. By employing a discrete model
P ∈ R
M×N
of the physical projection operator P
phys
, projection data p
0
= Pf
0
of the approximative result can be calculated. Now, the difference p
in
− p
0
can
be used to create an approximate correction image f
diff
= Q(p
in
− p
0
). Finally,
addition of f
0
and f
diff
yields a corrected result f
1
, which under certain circum-
stances contains less artifacts than f
0
. By repeating this procedure, a sequence of
gradually improved results f
0
, f
1
, ... is obtained.
In the field of CT, IFBP methods have successfully been applied to problems
where data are missing due to heavy objects [66] and due to an incomplete source
trajectory [123]. Danielsson and Magnusson [12] proposed to use this scheme to
suppress artifacts caused by non-exact reconstruction in helical cone-beam CT
with high cone angles.
1.4 Thesis outline
Chapter 2: Analytic reconstruction methods. Some basic concepts in CT-
reconstruction such as ramp filtering and backprojection are presented, followed
by a short description of the WFBP method.
Chapter 3: Iterative reconstruction methods. As an introduction we de-
scribe the well-known Landweber method [55], which is one of the simplest meth-
ods for iterative reconstruction, followed by examples and comments on various
parameters and alternative methods. In Section 3.3, we discuss the concepts of
basis and irradiation functions. This is followed by an introduction of regulariza-
6 Introduction
f
diff
f
k
p
k
+
+
+
Q
P
-

p
in
f (x, v, z)
P
phys
Figure 1.5: Iterative Filtered Backprojection (IFBP)
tion techniques for least square methods. The motivation behind this is that the
similarity, and sometimes even equivalence, between least squares methods and
IFBP methods suggests a way to introduce regularization with penalty terms for
the IFBP methods.
Chapter 4: Simulation and evaluation. Here, we describe how simulated pro-
jection data will be generated in the subsequent chapters. Geometrical parameters
and the different phantoms are explained and specified. The methods for assessing
spatial resolution and measuring noise and errors are also presented.
Chapter 5: Regularized Iterative Filtered Backprojection (RIWFBP).
This chapter starts with a description of the Regularized Iterative Weighted Filtered
Backprojection (RIWFBP) method, which is one of the main contributions in this
thesis. Besides a description and analysis of the iterative method, we discuss the
design of the regularization operator, and the possibility to iteratively reconstruct
a region of interest that is partially or completely contained by the examined
object. Subsequent sections comprise experimental studies of artifact suppression,
spatial resolution versus noise, behavior over multiple iterations, and local region
of interest reconstruction.
Chapter 6: Ordered Subsets IWFBP (OS-IWFBP). Ordered Subsets (OS)
is a well-known technique for reducing the number of required iterations. In this
chapter, we present the OS technique in general and show how it can be applied
to the RIWFBP method. Experiments with high cone angles (κ
max
= 4.8

and
κ
max
= 9.6

) show how OS-IWFBP outperform RIWFBP with respect to cone
artifact suppression in these cases.
Chapter 7: Relation between RIWFBP and least squares methods. Be-
cause of an advanced normalization procedure in the WFBP method, the RIWFBP
method is not easily expressed in terms of optimization theory. In this chapter,
a simplification of the normalization in the RIWFBP is presented, leading to the
1.5 Contributions and publications 7
Weighted Least Squares IFBP (WLS-IFBP) method. Evaluation by visual inspec-
tion, noise measurements, and root mean squared error measurements show that
WLS-IFBP produces images very similar to those of RIWFBP, but with a slightly
slower rate of convergence than for RIWFBP. In return, the WLS-IFBP method is
easier to analyze and improve using existing theory from the field of optimization
theory.
Chapter 8: Nonlinear regularization. First, we describe how nonlinear regu-
larization, i.e., non-quadratic potential functions, can be used with the RIWFBP
and WLS-IFBP. This is followed by experiments with WLS-IFBP and the q-
Generalized Gaussian Markov Random Field (q-GGMRF) potential function by
Thibault et al. [109]. Noise and spatial resolution measurements at different con-
trast levels show that nonlinearly regularized WLS-IFBP can be used to improve
high contrast spatial resolution and reduce artifacts while increasing the signal to
noise ratio for low-contrast regions.
Chapter 9: Data acquisition modeling. A careful choice of the forward projec-
tion operator P is important for obtaining good results in iterative reconstruction
methods. In the first part of this chapter, we investigate the effects of accurate
multiple ray models of the geometry during data acquisition. It is shown that
accurate modeling of source, detector, and gantry rotation can improve spatial
resolution, and reduce artifacts related to aliasing. The second part of this chap-
ter presents a new strip integrating projection operator that can reduce aliasing
related artifacts and improve spatial resolution.
Chapter 10: Implementation on Graphics Processing Units (GPU). The
forward projection and backprojection operations in CT reconstruction are com-
putationally very expensive. In this chapter, we describe our experiences in ac-
celerating these operations by implementing them on graphics hardware using the
NVIDIA Compute Unified Device Architecture (CUDA).
Chapter 11: Conclusions and future research. This chapter provides a
summary of conclusions and presents some questions remaining to be answered.
1.5 Contributions and publications
Much of the material in this thesis has been adapted from previous publications.
Below follows a brief summary of the contributions, and how they are related to
existing publications.
The presentation and evaluation of RIWFBP method in Chapter 5 is one of
the major contributions in the thesis. An earlier version without regularization or
evaluation of spatial resolution versus noise was first published in [103]. In [102,
ch. 5] and [105], regularization and spatial resolution versus noise were examined.
Further improvements of the method, as well as experiments with data from a
8 Introduction
clinical scanner were published in [106].
The WLS-IFBP method in Chapter 7 and nonlinear regularization in Chapter
8 were initially published in [107]. Since then, corrections and improvements
of the theory and reconstruction methods have been made. Furthermore, the
evaluation have been improved by including assessment of contrast dependent
spatial resolution and reconstruction of data from a clinical scanner.
The first part of Chapter 9 examines effects of more accurate forward projec-
tion models and effects of backprojector type, and has not yet been published.
This contribution provides insight in how accurate, but computationally expen-
sive, modeling can affect reconstruction results, and what one can expect from
good approximations of the models examined here.
The second part of Chapter 9 presents a strip integration technique for aliasing
suppressed forward projection and backprojection [104].
1.6 Notation
A selection of symbols and a list of acronyms are found in the appendices. Below,
we explain the basic mathematical conventions used in the thesis.
• Scalars and scalar-valued functions are denoted with italic letters, e.g. x,
f(x), and f(r).
• Vectors and matrices are denoted by bold lowercase and uppercase letters
respectively, e.g. r and P. The jth component of a vector is written as r
j
.
Correspondingly, an element on the ith row and jth column of a matrix P
is denoted by p
ij
. The superscript
T
is used for denoting the transpose of a
vector or a matrix.
• General linear operators are denoted by italic uppercase letters, e.g. P
phys
.
• Euclidean scalar products and norms are denoted by ', ` and | | re-
spectively. A subscript on these symbols means weighting according to
|x|
2
W
= 'x, x`
W
= x
T
Wx.
Chapter 2
Analytical reconstruction
methods
2.1 Two-dimensional image reconstruction
The mathematical problem of two-dimensional image reconstruction from line in-
tegrals was first solved by Johann Radon in 1917 [81]. Practical solutions suitable
for computer implementations have been an active research topic since early 1970s.
Several text books on the subject exist, see for instance the books by Herman [36],
Kak and Slaney [45] and Natterer [70]. Here we present the reconstruction prob-
lem in general followed by a closer look at the widely used filtered backprojection
(FBP) method by Bracewell and Riddle [6], and Ramachandran and Lakshmi-
narayanan [82]. The presentation is mainly inspired by the more extensive text by
Kak and Slaney [45], and works by Magnusson [64] and Turbell [110].
Consider Fig. 2.1 that illustrates an x-ray beam with initial photon intensity
I
0
traveling through an object described by the linear attenuation function f(x, y).
After traveling through the object, the remaining photon intensity I(t, θ) is given
by
I(t, θ) = I
0
e


Γ
t,θ
f(x,y)dl
. (2.1)
Taking the logarithm yields
p(t, θ) = −log

I(t, θ)
I
0

=

Γ
t,θ
f(x, y)dl. (2.2)
The linear operator { mapping f(x, y) to p(t, θ) is known as the Radon trans-
form. Any reconstruction method for two-dimensional parallel beam CT faces the
problem of inverting this transform.
In order to invert the Radon transform, the Fourier slice theorem is very useful.
It provides a connection between the Radon transform and the Fourier transform
of a two-dimensional function:
10 Analytical reconstruction methods

¦0 )=

-1
I
0
¦ v)al

v

I
0
v
Figure 2.1: Photon intensities of attenuated and non-attenuated x-ray beams
differ by a factor e


Γ
t,θ
f(x,y)dl
.

pil domin Fourier domin

¦ 0)=¦ )¦ 0)
¦ )¦¦)¦))
=0n/
0
Figure 2.2: Illustration of the Fourier slice theorem.
Theorem 1 (The Fourier slice theorem) The two-dimensional Fourier trans-
form Tf of a function f(x, y) along a radial line equals the one-dimensional
Fourier transform along the radial direction of the Radon transform:
(T
2
f)(ρ cos φ, ρ sin φ) = (T
t
({f))(ρ, θ) = (T
t
p)(ρ, θ), φ = θ −
π
2
. (2.3)
The theorem is illustrated in Fig. 2.2 and a formal proof is given by Kak and
Slaney in [45].
Theorem 1 can be applied directly to parallel projection data by computing the
two-dimensional Fourier transform of f(x, y) through a series of one-dimensional
Fourier transforms T
t
({f). Performing a final inverse two-dimensional Fourier
transform yields the reconstructed result. However, applying Theorem 1 directly
is associated with the problem of mapping radially sampled Fourier data to a
2.1 Two-dimensional image reconstruction 11
Cartesian grid. Even if solutions to this problem exist (see Magnusson [64] and
O’Sullivan [77]), these Direct Fourier Methods (DFMs) have been less popular
than filtered backprojection (FBP).
An FBP method implements the inversion formula
f(x, y) =

π
0
(p(, θ) ∗ g

)
. .. .
filtering
(xsin θ −y cos θ, θ)dθ
. .. .
backprojection
. (2.4)
As a first step, the projection data p = {f are filtered with a linear and spatially
invariant so called rampfilter g

in the radial (t) direction. This step is followed
by a backprojection, which consists of adding (in the case of a finite number of
angles) the filtered projection data to a digital image along the rays from which
the projection data has its origin. In terms of operator theory, backprojection
is the Hilbert adjoint operator {

to { [70]. Projection and backprojection are
illustrated in Fig. 2.3.
Fig. 2.3, inspired by Magnusson [64], also gives an intuitive explanation for
the Point Spread Function (PSF) of the combined operation projection followed
by backprojection. Assume that projections of a very small object have been
collected and are being backprojected as shown to the right in Fig. 2.3. Then,
the resulting image f
bp
(x, y) is a good approximation of the point spread function
that we search for. Clearly, this image will be non-zero not only on the support
of the original object, but also in the rest of the image. In fact, by observing
backprojected contributions to circles centered around the object of interest, we
may conclude that f
bp
(x, y) decays as 1/d, where d is the distance to the object.
Since the 2D Fourier transform of this function is 1/

u
2
+v
2
, the inverse filter
should look like

u
2
+v
2
in the Fourier domain. According to Theorem 1, this
filtering can be implemented as a one-dimensional filtering in the radial direction
of the projection data p(t, θ).
Equation (2.4) can be derived mathematically as follows.
f(x, y) =


−∞


−∞
(Tf)(u, v)e
i2π(xu+yv)
dudv
= (u = ρ cos φ, v = ρ sin φ, dudv = [ρ[dρdφ)
=

π/2
−π/2


−∞
(Tf)(ρ cos φ, ρ sin φ)
. .. .
=(F
t
p)(ρ,φ+π/2)
e
i2πρ(x cos φ+y sin φ)
[ρ[dρdφ
=

π
0


−∞
(T
t
p)(ρ, θ)[ρ[e
i2πρ(x sin θ−y cos θ)
dρdθ
=

π
0
(T
−1
t
(T
t
p Tg

))(xsin θ −y cos θ, θ)dθ
=

π
0
(p(, θ) ∗ g

)(xsin θ −y cos θ, θ)dθ (2.5)
12 Analytical reconstruction methods
Projection Backprojection
f ¦ v
¦ =¦ f ¦ ¦
f

¦ v=¦

¦ v
Figure 2.3: Illustration of projection followed by backprojection. Without any
filtering of projection data, this operation has a low-pass filtering effect on the
image (see text).
Here, g

is a generalized function satisfying Tg

= [ρ[. Because of its shape in
the Fourier domain, this filter is called a rampfilter.
In a practical implementation, all signals involved are digital. Therefore, the
rampfilter g

must be bandlimited and sampled before it can be used. From the
theory of sampled signals (see for instance Bracewell [5]), we know that sampling
with a distance Δ
t
is equivalent to convolution with the Shah function
1
III
Δ
−1
t
in the Fourier domain. Hence, the Fourier transform of the sampled rampfilter
is Δ
−1
t
-periodic. As shown in Fig. 2.4, one way to avoid discontinuities in this
periodic Fourier transform is to multiply Tg

with a rectangle of width Δ
−1
t
and
sample its inverse Fourier transform. This inverse Fourier transform is given by
g(t) = T
−1

Δ
−1
t
2

Π(ρ/Δ
−1
t
) −Λ(2ρ/Δ
−1
t
)

=
1

2
t
sinc(t/Δ
t
) −
1

2
t
sinc
2
(t/(2Δ
t
)). (2.6)
where Π and Λ are rectangle and triangle functions respectively, and
sinc(x)
sin(πx)
πx
. (2.7)
Sampling of g with sampling distance Δ
t
results in the sequence
g
s
[k] = g(kΔ
t
) =





1

2
t
k = 0
0 k even

1
k
2
π
2
Δ
2
t
k odd
. (2.8)
1
The Shah function III
Δ
is defined as an infinite sequence of Dirac impulses separated and
weighted with a distance Δ, i.e., III
Δ
(t) Δ


k=−∞
δ(t −kΔ).
2.2 Weighted filtered backprojection 13

))

1
2

1
2

Figure 2.4: By bandlimiting the rampfilter so that no frequencies higher than
Δ
−1
t
/2 exist, the Fourier transform of the bandlimited and sampled rampfilter
becomes continuous.
Rampfiltered projection data can now be calculated as
q[k, θ] = Δ
t

¸
l=−∞
g
s
[l]p[k −l, θ] (2.9)
where p[k, θ] are the original projection data for a certain angle θ. We are only
interested in the result of this convolution in the interval where p[k, θ] is non-zero.
Therefore, the summation in (2.9) can be made finite by truncating g
s
[k] to the
double length of this interval.
Discretization of the backprojection operation is done in the following way.
Suppose that projection data have been collected for the angles θ
0
, ..., θ
N
θ
−1
, and
radial coordinates t
0
, ..., t
N
t
−1
. For each angle θ
l
, a continuous projection q(t, θ
l
)
is created by means of linear interpolation, i.e.
q(t, θ
l
) =
N
t
−1
¸
k=0
q[k, θ
l

t −t
k
Δ
t

. (2.10)
The final image can now be created by approximating the integration over angles
with a summation, resulting in
f
rec
(x, y) =
π
N
θ
N
θ
−1
¸
l=0
q(xsin θ
l
−y cos θ
l
, θ
l
). (2.11)
2.2 Weighted filtered backprojection
The Weighted Filtered Backprojection method (WFBP) by Stierstorfer et al. [100],
is an approximative reconstruction method for helical cone-beam CT, and one
of the main components in iterative methods studied in later chapters. In the
following, we give a presentation of the WFBP method, mainly based on [100].
The three main steps are
14 Analytical reconstruction methods
Physical detector
x
y
Virtual detectors
x
y
x
z
x
z
Figure 2.5: Left: the original cone-beam geometry. Right: the semi-parallel
geometry obtained after rebinning. Because of the specific shape formed by the
rays, this geometry is also called the wedge geometry [111].
1) Rebinning of cone-beam data to semi-parallel data (Fig. 2.5),
2) Ramp-filtering of rebinned projection data, and
3) Normalized backprojection.
To understand these steps, the geometries of Fig. 2.6 and 2.7 must be under-
stood. Given a certain source angle α, the location of the source is determined
by
x
S
(α) = R
S
cos α,
y
S
(α) = R
S
sin α, and (2.12)
z
S
(α) =


,
2.2 Weighted filtered backprojection 15

X-ray source
Figure 2.6: Scanning geometry in the xy-plane.

P
X-ray source ( ÷ 0)
Detector

max
1
0
-1

D
Figure 2.7: Scanning geometry in the yz-plane.
where R
S
is the distance between the source and the isocenter, and P is the
patient table feed, or equivalently, the pitch of the helix. An arbitrary ray is
determined by the angles α, β and a continuous variable q ∈ [−1, 1] describing the
slope of the ray, where q = ±1 corresponds to rays hitting the top and bottom of
the detector respectively (see Fig. 2.7). Given these parameters, arbitrary points
along the ray can be calculated by
x
α,β,q
(l) = x
S
(α) −l cos(α +β),
y
α,β,q
(l) = y
S
(α) −l sin(α +β), and (2.13)
z
α,β,q
(l) = z
S
(α) +ql tan κ
max
.
Here, l determines a particular point on the ray. For instance, l = 0 points out the
source position, and l = R
S
+R
D
points out the position of the detector element.
16 Analytical reconstruction methods
Fig. 2.6 shows how an arbitrary ray is uniquely determined in the xy-plane,
either by specifying α and β or by specifying the angle θ and the orthogonal
distance t to the isocenter. Given α and β, the parameters θ and t can be calculated
from
θ = α +β, t = R
S
sin β. (2.14)
The other way around, if θ and t are known, α and β can be obtained by using
the relation
α = θ −β, β = sin
−1

t
R
S

. (2.15)
The procedure of translating 2D projection data parametrized with (α, β) to par-
allel data (θ, t) is referred to as rebinning. In the first step of the WFBP method,
this is done independently for each detector row. The resulting geometry is illus-
trated in the right image in Fig. 2.5. Projected onto the xy-plane, the rebinned
projections look perfectly parallel. However, the rays diverge in the z-direction
and the rebinned projections are therefore called semi-parallel projections.
As a second step in the WFBP method, rampfiltered projections p
conv
(θ, t, q)
are obtained from p(θ, t, q) by applying the rampfilter from the previous section.
Similarly to the rebinning, rampfiltering is done on each detector row indepen-
dently, i.e., data are not mixed between detector rows. The rampfilter can be
combined with a low-pass filter m(t) to shape the frequency response of the CT
system. Commercial systems are shipped with a variety of such filters, designed
for different clinical applications. Throughout this thesis, we refer to these filters
as resolution modifying filters.
Similarly to two-dimensional backprojection, the WFBP backprojection con-
structs a voxel volume as a sum of partial backprojections
f
rec
(x, y, z) =
N
π
−1
¸
n=0
f
˜
θ
n
(x, y, z) (2.16)
over different angles
˜
θ
n
, where N
π
is the number of projection angles per half turn.
However, in contrast to (2.11), each term in this equation contain contributions
from several θ
n
-angles, namely θ
n
=
˜
θ
n
+kπ, k = 0, 1, .... The reason for dividing
the backprojection in such partial backprojections is illustrated in Fig. 2.8. This
figure shows how a certain voxel may receive one, two or three contributions from
one
˜
θ
n
-angle depending on the location of the voxel. Because of this, and the
weighting described in the previous paragraph, normalization must be performed
for each individual
˜
θ
n
-angle. Below, we describe how the contribution from one
such angle is calculated. A more detailed description of the backprojection is given
in Section 10.5.1.
Given a point (x, y, z) in which we want to know the backprojected values,
a first step is to calculate the projection (
ˆ
t
k
, ˆ q
k
) of this point onto the virtual
2.2 Weighted filtered backprojection 17

1
- 1

1

2

3
Figure 2.8: Example of how different voxels receive different numbers of contri-
butions during backprojection from a certain angle
˜
θ. Voxels located in the lightly
shaded area receive a contribution from one focus position. In the medium and
dark shaded areas, voxels receive two and three contributions respectively.
detector plane illustrated in Fig. 2.5. This is given by
ˆ
t
k
= xsin(
˜
θ
n
+kπ) −y cos(
˜
θ
n
+kπ),
ˆ q
k
=
z −P

˜
θ
n
+kπ −sin
−1
ˆ
t
k
R
S

/2π

R
2
S

ˆ
t
2
k
−xcos(
˜
θ
n
+kπ) −y sin(
˜
θ
n
+kπ)

R
S
h/2
. (2.17)
(2.18)
Knowing
ˆ
t
k
and ˆ q
k
, a weighted and normalized partial contribution can be calcu-
lated as
f
˜
θ
n
(x, y, z) =
1
N
k
¸
k
p(
˜
θ
n
+kπ,
ˆ
t
k
, ˆ q
k
), (2.19)
where the value of p(
˜
θ
n
+kπ,
ˆ
t
k
, ˆ q
k
) is obtained by linear interpolation with respect
to
ˆ
t
k
and ˆ q
k
, and N
k
is the number of non-zero contributions in the sum over k.
As shown in Fig. 2.9 a), the sharp transitions between regions receiving one,
two, or more contributions, give rise to severe non-local artifacts. To alleviate
these artifacts, the backprojection formula in (2.19) is modified into
f
˜
θ
n
(x, y, z) =
1
¸
k
W
Q
(ˆ q
k
)
¸
k
W
Q
(ˆ q
k
)p(
˜
θ
n
+kπ,
ˆ
t
k
, ˆ q
k
), (2.20)
where the weighting function W
Q
(q) is given by given by
W
Q
(q) =

1 [q[ < Q

cos

π
2
|q|−Q
1−Q

2
Q ≤ [q[ < 1
. (2.21)
This function is illustrated in Fig. 2.10 for different choices of Q. In Fig. 2.9 b),
showing a WFBP reconstruction with Q = 0.7, we see that the artifacts caused
by the sharp transitions are almost perfectly eliminated.
18 Analytical reconstruction methods
a) Q = 1.0 b) Q = 0.7
Figure 2.9: WFBP reconstructions of the Turbell Clock phantom [110] with
Q = 1.0 and Q = 0.7 respectively. Greyscale window ±20HU.
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
Normalized row coordinate q ∈ (−1, 1)
W
e
i
g
h
t
i
n
g
Q = 0.0 Q = 0.3 Q = 0.7 Q = 1.0
Figure 2.10: The weighting function W
Q
(q) for Q equal to 0.0, 0.3, 0.7, and 1.0.
Chapter 3
Iterative reconstruction
methods
3.1 A basic example. Application overview.
As mentioned in Chapter 1, iterative methods generate a sequence of estimates that
eventually should converge towards a good solution. How this works in principle
is illustrated here with the Landweber method, originally developed for iteratively
solving Fredholm equations of the first kind [55].
Let the estimates of linear x-ray attenuation coefficients be represented by
vectors f
k
∈ R
N
, and let p
in
∈ R
M
represent input projection data. Here, k refers
to a certain iteration and each component f
j
in f
k
=

f
1
f
N

T
represents a
certain voxel (see (3.8)). A detailed discussion of the relation between the vector
f
k
and corresponding image by means of basis functions is found in the following
section. The updating formula for the Landweber method is given by
f
k+1
= f
k
−α
k
P
T
(Pf
k
−p
in
) (3.1)
and is illustrated in Fig. 3.1. The matrix P ∈ R
M×N
represents a linear model
of the physical acquisition process. Thus, P maps the estimate f
k
onto Pf
k
, the
projection of the estimate. The difference Pf
k
−p
in
is then used to create a new
estimate by backprojection with the adjoint operator P
T
∈ R
N×M
, the result of
which is multiplied with the step length α
k
, and finally subtracted from the old
estimate.
By keeping α
k
constant with respect to k and letting Δ = I − αP
T
P, we see
that
f
k+1
= Δf
k
+αP
T
p
in
= Δ
2
f
k−1
+ (Δ+I)αP
T
p
in
= Δ
k+1
f
0
+ (Δ
k
+ +Δ+I)αP
T
p
in
. (3.2)
20 Iterative reconstruction methods
f
diff
f
k
p
k
+
+
+

-

p
in
Figure 3.1: Illustration of the Landweber method. In order to calculate a new
estimate f
k+1
, projection data Pf
k
are first calculated. These are then subtracted
from the input data and a difference image (gradient) is obtained by backprojection
with the adjoint operator P. This difference image is finally multiplied with a step
length α and is added to the old estimate f
k
.
Therefore, if P
T
P is positive definite and α is chosen so that the spectral radius
of Δ is less than one, then the first term become zero and the expression (Δ
k
+
+Δ+I) converges to (I −Δ)
−1
. This leads us to the fixed point
f

= lim
k→∞
f
k
= (I −Δ)
−1
αP
T
p
in
= (P
T
P)
−1
P
T
p
in
. (3.3)
In the updating formula (3.1), let
P
T
(Pf −p
in
) = ∇z(f ). (3.4)
Since z(f ) can be identified as
z(f ) =
1
2
|Pf −p
in
|
2
2
, (3.5)
the method (3.1) is recognized as a steepest descent method for minimizing z(f )
[71].
Compared to analytical reconstruction, most iterative methods are slow. How-
ever, there are cases where this drawback is outweighed by the relative ease by
which the iterative methods handle the modeling problem. Any analytical method
faces the problem of modeling the inverse of every physical phenomenon involved
in the data capture. In contrast, iterative methods require an emphasis primarily
on modeling the forward versions of all such processes.
Some examples where iterative methods have been found worthwhile to explore
are listed below.
• Redundant data. In helical cone-beam CT, at this time of writing and to
the best of our knowledge, there is no exact method able to exploit redundant
data for better Signal-to-Noise Ratio (SNR), a fact which has forced the CT-
industry to rely on non-exact methods. Iterative methods can simultaneously
produce accurate images and make use of all available data.
3.2 Different types of iterative methods 21
• Better noise models. Recently, several authors have shown how improved
statistical models of input data can improve the SNR in helical cone-beam
CT [73, 16, 121, 127, 109, 98].
• Incomplete data. One example of how to suppress artifacts due to missing
data in circular cone-beam CT is given by Zeng et al. [123]. Here the
Feldkamp method is applied iteratively to reduce errors due to the incomplete
circular source trajectory used for data capture.
• Resolution recovery. In studies by Nuyts et al. [73] and Thibault et
al. [109], iterative methods were shown to increase the spatial resolution in
comparison to filtered backprojection methods.
• Beam hardening correction. De Man et al. [21] and Elbakri and Fessler
[23] have shown that by simulating multi energetic x-rays in the acquisition
model, beam hardening artifacts can be significantly reduced.
• Metal artifact reduction. De Man et al. [20] have shown that iterative
methods can be used to reduce artifacts caused by the presence of strongly
attenuating objects.
In the following sections, we give a brief presentation of different types of iter-
ative reconstruction methods, as well as operations and techniques that frequently
appear in iterative reconstruction methods.
3.2 Different types of iterative methods
Most iterative reconstruction methods are based on the minimization or maxi-
mization of a cost function z(f , p
in
). Such a cost function is typically expressed
as
z(f , p
in
) = z
D
(f , p
in
) +z
R
(f ), (3.6)
where z
D
measures the difference between the projected volume Pf and input
data p
in
, and z
R
is a regularization term enforcing some kind of smoothness on
the reconstruction result f . If the cost function is convex, any globally convergent
optimization method will reach the optimum. However, from a practical point
of view, the choice of optimization method is critical, since iterations are often
stopped before the fixed point has been reached.
Xu et al. [117] have divided iterative methods into three categories: conven-
tional algebraic reconstruction techniques, iterative statistical methods, and IFBP
methods. Below follows a short description of each of these categories.
Conventional algebraic reconstruction techniques include Algebraic Re-
construction Technique (ART) [31], Simultaneous Algebraic Reconstruction Tech-
nique (SART) [1], Simultaneous Iterative Reconstruction Technique (SIRT) [30]
22 Iterative reconstruction methods
and the Landweber method presented above, among others. It was shown by Jiang
and Wang [40] that under certain circumstances, all of them minimize weighted
square norms similar to (3.5).
Statistical image reconstruction methods are based on the Poisson model
for the photon intensity. This is used for deriving the conditional likelihood
P(p
in
[f ) for the acquired input data given a certain vector f of linear attenu-
ation coefficients. Bayes’ rule gives an expression for the reciprocal conditional
likelihood
P(f [p
in
) =
P(p
in
[f )P(f )
P(p
in
)
. (3.7)
The aim of statistical methods is to maximize log P(f [p
in
). If P(f ) is assumed to be
constant, the method is called a Maximum Likelihood (ML) method. Otherwise,
the method is called a Maximum a Posteriori (MAP) method. Many methods for
maximizing the ML and MAP functions exist. Examples of these are Maximum
Likelihood Expectation Maximization (MLEM) [56], the convex method [57], and
the ordered subsets convex (OSC) method [47]. The ML and MAP methods
are known to produce images with better signal to noise ratio, and many recent
developments toward faster methods make these methods promising. Recently,
interesting results have also been presented from using the simplified cost function
obtained by approximating log P(p
in
[f ) with its second order Taylor expansion
[50, 109].
Iterative Filtered Backprojection (IFBP) methods are based on iterative
application of FBP methods. The simplest IFBP method is obtained by substitut-
ing P
T
with Q in the Landweber method, where Q is a FBP method. Since the
FBP method is used in each iteration, certain artifacts caused by non-exactness
of Q are very rapidly suppressed. Therefore, for the purpose of suppressing such
artifacts, IFBP methods are usually much faster than other iterative methods.
For attenuation correction in Single Photon Emission Computed Tomography
(SPECT), IFBP was proposed by Chang [8], and Walters et al. [113], while Medoff
et al. [66] used IFBP for improved reconstruction from limited angles. Censor et
al. [7] presented and analyzed the general framework of Iterative Data Refinement
(IDR), that includes IFBP as a special case. It was later shown by Xu et al. [117]
that for emission tomography, IFBP methods have computational advantages over
the conventional and statistical methods. Furthermore, the SNR in relation to
Contrast Recovery Coefficients (CRC) of IFBP methods were comparable to those
of regular FBP. Similar results for transmission tomography were later presented
by Nuyts et al. [73]. Recently, Zeng et al. [123] employed an IFBP method
for reduction of circular trajectory artifacts in the Feldkamp algorithm. Further
acceleration of IFBP methods have been suggested by Riddell et al. [83], and
Delaney and Bresler [22]. IFBP methods, which are the main topic of this thesis,
will be presented in more detail in Chapter 5, where we introduce the Regularized
Iterative Weighted Filtered Backprojection (RIWFBP) method.
3.3 The projection operator 23
b) Bilinear basis Iunction

()÷ (/

)(/

)
a) Image coeIIicients

1

6

3

4

7

10

13

16

5

8

9

11

12

14

15

17

18

2

c) Interpolated image

¦

1
18

¦-

Figure 3.2: Illustration of how a two-dimensional array of image coefficients are
mapped onto a continuous function by interpolation with a bilinear basis function.
The sampling distance in the y-direction is assumed to be equal to the sampling
distance in the x-direction.
3.3 The projection operator
3.3.1 Matrix representation
The following derivation of a matrix formulation of the projection operation P has
mainly been inspired by a paper on local basis functions by Hanson and Wecksung
[33]. A continuous three-dimensional image, i.e., a not necessarily continuous
function that maps R
3
into R, can be represented as a linear combination of basis
functions as
f
c
(r) =
N
¸
j=1
f
j
b
j
(r), (3.8)
where r ∈ R
3
is a coordinate vector and b
j
: R
3
→ R are the basis functions (for
the 2D case, see Fig. 3.2). Usually, these basis functions are translated copies of
a single basis function b(r) so that (3.8) becomes
f
c
(r) =
N
¸
j=1
f
j
b(r −r
j
). (3.9)
24 Iterative reconstruction methods
t
i
a) The Iunction f
c
() b) The irradiation Iunction w
i
()
c) The product w
i
() f
c
()
t
i
t
i

p
i

2
w
i
f
c
axav
Figure 3.3: Illustration of a linear model for projection generation in the case of
two dimensions. The image f
c
(r) is first multiplied with an irradiation function
specifying the photon flux in different areas of the image. The result is then
integrated to get the final result p
i
=

R
2
w
i
(r)f
c
(r)dr.
where each r
j
corresponds to the center of a certain voxel.
Fig. 3.3 shows a linear model for the acquisition process in the case of two-
dimensional functions. In the three-dimensional case, if nonlinear effects are ne-
glected, the contribution from a continuous three-dimensional image f
c
: R
3
→R
to a certain detector reading i can be calculated as
p
i
=

R
3
w
i
(r)f
c
(r)dr. (3.10)
The three-dimensional functions w
i
(r) will throughout this thesis be called irra-
diation functions. They constitute a model for the source and detector system
and contain values of the photon flux for a certain reading i in a certain point r.
The simplest and most common choice of functions w
i
are Dirac lines, resulting in
pure line integral contributions. More elaborate functions such as strips or sums
of several Dirac lines can be used to create more realistic models of the acquisition
process (to be discussed in Chapter 9).
Insertion of (3.9) into (3.10) yields
p
i
=

R
3
w
i
(r)
N
¸
j=1
f
j
b(r −r
j
)dr =
N
¸
j=1
f
j

R
3
w
i
(r)b(r −r
j
)dr (3.11)
From this expression, we see that the projection data p =

p
1
p
M

T
can
3.3 The projection operator 25
be calculated by a matrix multiplication p = Pf where f =

f
1
f
N

T
and
the components of P ∈ R
M×N
are given by
p
ij
=

R
3
w
i
(r)b(r −r
j
)dr. (3.12)
For simple irradiation functions such as Dirac lines and strips, the above computa-
tion can be efficiently implemented as a simple lookup in a table of precomputed
values (see for instance [124] and [69]). Clearly, if w
i
(r) is a Dirac line, (3.12)
yields a line integral taken through the basis function. This observation takes us
to the important notion of footprints. Let s ∈ S
2
⊂ R
3
be a direction vector and
let t ∈ s

⊂ R
3
describe a displacement in the orthogonal complement of s. Then
the footprint g(s, t) of the basis function b(r) is defined by
g(s, t) =

R
b(t +ls)dl. (3.13)
According to the Fourier slice theorem, the Fourier transform of g(s, t) with respect
to t equals the Fourier transform of b(r) on s

. Thus, the footprint can be used to
study Fourier domain properties of a basis function. Later in this section, examples
will be given of basis functions that are defined by specifying their footprints rather
than directly specifying b(r). In cases where the footprint is invariant to changes
in direction s and the irradiation functions w
i
(r) are Dirac lines, footprint tables
can be used for efficient implementations of P [69].
3.3.2 Basis functions
The choice of basis function b(r) as well as the irradiation functions w
i
(r) affect
the result of an iterative method. A good basis function should (i) be able to ac-
curately represent a constant function, (ii) allow for cost-effective implementation
of forward projection and backprojection operations, and (iii) contain a minimal
amount of frequency components higher than the Nyquist frequency, since these
give rise to aliasing artifacts. Below, we discuss the first of these properties in
more detail.
Requirement (i) is easy to understand and verify for basis functions which are
separable as b(x, y) = b
x
(x)b
y
(y). Examples of such basis functions satisfying this
requirement are the Siddon basis function (nearest neighbor interpolation) [92],
and the K¨ ohler basis function (trilinear interpolation) [51]. Rotationally symmetric
basis functions and basis functions implicitly defined by their footprints are harder.
Ohyama et al. [76] and Danielsson and Magnusson [12] solved the verification
problem by studying the Fourier transforms of the involved functions. Assuming
that the sampling distance equals Δ
x
, the Fourier transform of the repeated basis
functions is given by

T

1
Δ
x
III
Δ
x
∗b

(u) = (III
Δ
−1
x
Tb)(u), (3.14)
26 Iterative reconstruction methods
where III
Δ
x
(t) = Δ
x
¸

k=−∞
δ(t − kΔ
x
) (see p. 12). To be able to represent
the underlying constant function, the basis function (to be convolved with the
samples) must fulfill the following criterion. Its Fourier transform Tb must be
close to zero in all points where the impulses of III
Δ
−1
x
are located, except in the
origin. Otherwise, not only a DC-component will appear in the result, but also
false high frequency components. This aliasing of neighboring DC-components
was named “DC-aliasing” by Danielsson and Magnusson [12].
A simple example of a basis function that violates this requirement is illustrated
in Fig. 3.4 and 3.5. This basis function, which occurs as a special case of a
projection method proposed by Peters [78], is defined by its rotationally symmetric
footprint
g(s, t) =

Δ
x
−[t[ , [t[ ≤ Δ
x
0 , otherwise
. (3.15)
The Fourier transform of this basis function is given by
(Tb)(u, v) = Δ
2
x
sinc
2

x

u
2
+v
2
), (3.16)
and is exactly zero in the impulse points of III
Δ
−1
x
along a line aligned to one of the
main axes. However, if the Fourier transform is considered along a line tilted 45

away from one of the main axes, it is clear that there are non-zero contributions
at the impulse points of III
Δ
−1
x
. Thus, false frequency components appear when
this basis function is used for approximating a constant function. The forward
projection method by Joseph [41] overcomes this problem by scaling the footprint
depending on the projection angle. This method is further discussed in Section
3.3.4.
One well-known example of rotationally symmetric basis functions satisfying (i)
and (ii) above is the class of generalized Kaiser-Bessel basis functions (also known
as blobs) proposed by Lewitt [60, 61]. These functions are effectively bandlimited,
i.e., they are close to zero outside the Nyquist frequency of the reconstruction
volume. Another rotationally symmetric basis function based on the SinCot filter
by Magnusson [64] has been presented and evaluated in [12] and [101].
3.3.3 Irradiation functions
The most common type of irradiation function is a Dirac line transforming the
volume integral in (3.12) into a line integral along the line corresponding to mea-
surement i. There are at least two reasons to consider other types of irradiation
functions.
1) In order to capture all information delivered by the input data, it is desirable
to use a relatively high voxel density in comparison to the ray density. During
projection generation, such a relation between the detector and voxel density
may give rise to aliasing. This is realized by first studying the continuous
3.3 The projection operator 27
1

1

Zero crossings oI
) )=

2
sinc
2

2

2
)
Figure 3.4: Illustration of the periodic Fourier transform of a sampled image and
zero crossings for the function Δ
2
x
sinc
2

x

u
2
+v
2
).
projection of the continuous image obtained by convolution by the basis
function. The next step is to see whether the assumptions of the sampling
theorem [5] are violated during sampling of this continuous projection. If
the x-ray detector is configured so that the rays are relatively sparse in
comparison to the voxel density, aliasing may occur during projection. As
illustrated in Fig. 3.6, this type of aliasing will occur for a wide range of
sampling densities Δ
z
, for diverging rays appearing in the z-direction after
semi-parallel rebinning in WFBP.
Using an irradiation function that consists either of several Dirac lines or a
strip is similar to convolving the continuous projection with a certain low-
pass filter prior to sampling. Therefore, such irradiation functions can be
used for suppressing aliasing in the projection generation process.
2) The line integral model of the acquisition process is rather crude since the
28 Iterative reconstruction methods
1

2

1

2

1

2

1

2

2

2

2
sc
2


45º

2

2

RemagDC-compoet
causgalasgartIacts
PerIectsuppressooI
repeatedDCcompoets
Figure 3.5: Two intersections through the Fourier transform shown in Fig. 3.4.
Since Tb is zero for ρ =

u
2
+v
2
= ..., −Δ
−1
x
, 0, Δ
−1
x
, ..., the repeated DC-
components are perfectly eliminated for 0

. However, this is not true in the 45

direction where false high frequency components are introduced.
3.3 The projection operator 29
a) i Irequencies located close to te Iocus

2

b) i Irequencies located Iar Irom te Iocus

2

÷~ no aliasin

2

÷~ aliasin
Detector
Detector

Figure 3.6: An illustration of how aliasing occurs during projection generation.
High frequencies located close to the focus are projected to relatively low frequen-
cies on the detector, while high frequencies located far from the focus are projected
to relatively high frequencies on the detector.
gantry rotation and sizes of focus and detector elements are neglected. A
better model would take into account the smoothing caused by these ele-
ments.
Various examples of irradiation functions other than Dirac lines exist in the lit-
erature. In the papers by Hanson and Wecksung [33], and Ohyama et al. [76], strip
integrals (see Fig. 3.7) were used in combination with the square basis function in
two dimensions. In both papers, better results were achieved with strips than with
Dirac lines. The problem of aliasing was pointed out by Mueller et al. [68] and a
space variant scaling of the basis functions (in this case blobs) was suggested to
suppress aliasing artifacts. Later, Ziegler et al. showed that suppression of these
artifacts can be done by using divergent strip irradiation functions in combination
with blob basis functions [124]. De Man and Basu [17] proposed a technique for
generating projections that resembles the Joseph interpolation but instead of a
triangular interpolation function their interpolation employs two convolved rect-
angle functions with different widths. These widths were chosen to be the image
30 Iterative reconstruction methods
Focus Dirac line Detector element
Divergent strip
Non-divergent strip
Collection oI Dirac lines
Figure 3.7: Examples of different irradiation functions.
sampling distance and the spatial dependent ray distance respectively. Thus, in
terms of basis and irradiation functions, the first rectangle would correspond to
the basis function and the second to the irradiation function. We believe that this
idea should be highly beneficial to apply in the divergent projection field in Fig.
3.6.
In Chapter 9, the effect of using several Dirac lines per reading (see Fig. 3.7) for
anti-aliasing and more accurate modeling of the acquisition process is investigated.
3.3.4 The Joseph forward projection method
Unless otherwise noted, in the subsequent implementations and experiments we
have used the forward projection method proposed by Joseph [41]. This method
has the advantages of being simple to implement, computationally efficient, and
producing less aliasing artifacts than the method proposed by Siddon [92]. Here,
we give a brief description of the method in two dimensions. An extension to three
dimensions is straightforward but more difficult to illustrate.
Consider Fig. 3.8. Suppose that the line integral through an image along
a line Γ is to be calculated. The first step in the Joseph method is to divide
the image into a set of lines ¦L
i
¦
N
x
i=1
along the coordinate axis that is maximally
orthogonal to the line Γ. The next step is to use linear interpolation to calculate
contributions from each intersection between Γ and the lines ¦L
i
¦
N
x
i=1
. Finally, a
length correction is made by multiplying the sum of contributions with Δ
x
/ cos(γ),
where γ is the smallest angle between Γ and the coordinate axis perpendicular to
the lines ¦L
i
¦
N
x
i=1
.
The Joseph method does not suffer from DC-aliasing. Danielsson and Mag-
3.3 The projection operator 31

2

3

/ 2

/ 2

3

/ 2

/ 2
2

/ 3

I

Figure 3.8: Illustration of the Joseph projection method. Depending on the slope
of the line integral to be calculated, the image is divided into a set of lines such
that the angle γ does not exceed 45

. At each intersection between the line and
this set of lines, a contribution is calculated by means of linear interpolation. As
a final step, the sum of contributions is multiplied with Δ
x
/ cos(γ). This results
in a triangular footprint g(s, t) with a width that varies between 2Δ
x
and

x

2
.
nusson [12, 13] showed this by studying the footprint
g(s, t) =

1
cos(γ(s))

Δ
x

|t|
cos(γ(s))

[t[ ≤ Δ
x
cos(γ(s))
0 otherwise
(3.17)
where γ(s) is the angle in Fig. 3.8. Except for the compression factor cos(γ(s)),
this expression is identical to (3.15). By reasoning in the same way as was previ-
ously done for (3.15), we see that since the footprint is compressed with a factor
1

2
in the 45

-direction, the Fourier transform is stretched with a factor

2. There-
fore, all non-central repetitions of the DC-component are eliminated also in the
45

-directions, and in the directions arctan(±1/2) as well.
One problem with the Joseph method is that the footprint violates the Helgason
Ludwig conditions (see [70], Theorem 4.1)
1
. Therefore, there is no well-behaved
basis function (possible to approximate with a function in the Schwartz space) that
corresponds to this footprint. In cases where a basis function interpretation of the
Joseph method is needed, it can be useful to imagine a trilinear basis function
sheared as illustrated in Fig. 3.9 [110].
1
This fact was pointed out to the author by Dr. Michel Defrise.
32 Iterative reconstruction methods


a) b¦ . =0
°
b) b¦ . =20
°
Figure 3.9: Illustration of the Joseph basis function for different angles. The
figure is based on a similar figure in [110].
3.4 Regularization
A common problem with iterative image reconstruction methods is that the iter-
ates tend to become distorted by high frequency artifacts and noise as the number
of iterations grows. This is caused by (i) ill-posedness of the tomography problem
2
and (ii) insufficient restrictions on the space of possible solutions.
In the context of tomographic reconstruction, ill-posedness means that the pro-
jection operator P
phys
suppresses high frequencies rather efficiently, i.e., projection
data p
in
contain only small amounts of high frequencies. Therefore, if white noise
is added to the projection data, high frequency components of the projection data
will quickly be dominated by noise. Naively applying the Landweber method to
find the pseudo inverse of P as described in Section 3.1 leads to strong amplifica-
tion of these noisy components and unusable reconstruction results. A presentation
and discussion of the general problem of ill-posedness and the Fredholm equation
of the first kind is found in [32].
The problem with insufficient restrictions on the space of possible solutions
is illustrated in Fig. 3.10. The space of functions that maps the real line to
itself is too large and contains many solutions that we would consider incorrect
or noisy. Therefore, constraints must be added to restrict the solution space to
certain well-behaved functions. This is called regularization. In the context of
image reconstruction, at least four different regularization techniques appear in
the literature.
In [60], Lewitt propose to use rotationally symmetric generalized Kaiser-Bessel
basis functions (also known as “blobs”) for iterative reconstruction. The amplitude
2
A problem is said to be ill-posed if its solution is not unique, or if the solution does not
continuously depend on the input data, i.e., if a small perturbation of input data may lead to a
large perturbation of the solution [32].
3.4 Regularization 33
t
f
1
¦t
f

¦t
Figure 3.10: A simple example of fitting a function f : R → R to a number
of measurements. The filled circles illustrate the measured points and the curves
illustrate two possible functions f
1
and f
2
that interpolate the measured data.
spectrum of these functions is very close to zero outside the Nyquist frequency of
the image. In this way, the basis function is used for controlling the frequency
content of the reconstructed image. This kind of regularization has successfully
been applied to iterative image reconstruction in CT by Ziegler et al. [126].
Snyder and Miller [95], and Snyder et al. [96] proposed the use of sieves to
suppress the enhancement of noise as the number of iteration is increased. This
means that the iterative method is designed so that the result will lie in the so
called sieve
S =

f(r) : f(r) =

R
3
K(r −r

)f

(r

)dr

, K > 0, f

> 0

(3.18)
where K(r) is called the sieve kernel and is usually a smooth function, for instance
a Gaussian. The iterative method is designed to estimate f

by introducing a
convolution with the sieve kernel before projection as well as after backprojection
in each iteration. As a final step (after convergence), f

is convolved with the
sieve kernel to get f. Note that the continuous formulation (3.18) leaves room
for principally different implementations. For instance, f

can be identified as
a sequence of weighted Dirac impulses corresponding to the coefficients of the f
vector, and K can be identified as the basis functions. Then, sieve regularization
would be identical to regularization with a smooth basis function. Another, per-
haps more common interpretation is to design the iterative method so that the
resulting coefficient vector f lie in the sieve
S
discrete
=
¸
f ∈ R
N
: f = Kf

, K > 0, f

∈ R
N+
¸
(3.19)
where K correspond to a convolution with a discrete sieve kernel.
Another popular method for avoiding enhancement of noise is to stop the
method before the fixed point has been reached (see for instance Veklerov and
Llacer [112] and Hebert et al. [34]). The practical problem with this method is to
choose a good stopping criterion. This choice must be done with respect to the
particular optimization method being used.
Stopping an iterative method before convergence is theoretically unsatisfactory,
since this means that a sub-optimal point is preferred over the optimal solution
34 Iterative reconstruction methods
to the problem as it has been formulated mathematically. The conclusion is that
the mathematical problem formulation lacks some feasibility parameter not yet
accounted for. One way to improve the problem formulation is to use a penalty
term z
R
(f ) in the cost function (the second term in (3.6)). This term is designed
so that certain structures and properties considered likely to appear are penalized
less than other less likely structures.
One widely used choice of penalty function is based on the theory of Markov
Random Fields (MRF), or equivalently, Gibbs priors [63]. These penalty functions
have the general form
z
R
(f ) = β
N
¸
i=1
N
¸
j=1
d
ij
V (f
i
−f
j
), (3.20)
where β is called regularization parameter or hyperparameter, and d
ij
are usually
the inverse distances between voxels i and j within some small neighborhood,
e.g. of size 3 3 3. The function V is called potential function, and determines
how different contrast levels should be penalized. Regularization with this type of
penalty functions is a major topic of this thesis, and will be further discussed in
Chapter 5, 7 and 8.
Another penalty function that recently has appeared in the CT literature is
the Total Variation (TV) norm
z
R
(f ) = β|f |
TV
= β
N
¸
i=1

(∇
x
f )
2
i
+ (∇
y
f )
2
i
+ (∇
z
f )
2
i
, (3.21)
where ∇
x
, ∇
y
, and ∇
z
correspond to differentiations in the x-, y-, and z-directions
respectively. Some results have indicated that this penalty term can be useful in
reconstructions from projection data that have been very sparsely sampled along
the angular direction [93, 9].
Chapter 4
Simulation and evaluation
4.1 Generation of input projection data
Noise-free input data for the experiments in the following chapters were generated
with the CTSIM/DRASIM program package (Siemens Healthcare, Forchheim).
This program produces projection data in the form of line integrals. The main
input consists of a mathematical phantom f
pha
(r), and parameters describing the
scanning geometry. To simulate the finite extension of the source and the detec-
tor, the program divides these into N
SW
N
SL
and N
DW
N
DH
equidistantly
sampled points respectively (see Fig. 4.1), and calculates a mean value of several
line integrals instead of just one single line integral. To make the physical model
even more realistic, modeling of the gantry rotation has been added by calcu-
lating each angular measurement as the mean of N
GR
sub-measurements. Thus,
CTSIM/DRASIM in combination with the gantry rotation model calculates the
contributions after taking logarithms as
p
i
= −log


1
N
N
¸
j=1
exp

Γ
i,j
f
pha
(r)dl



(4.1)
where
N = N
SW
N
SL
N
DW
N
DH
N
GR
. (4.2)
As shown in Table 4.1, typically N = 3
5
= 243, which makes the computation
of these simulated projection data very expensive. However, since the same input
data can be used for several different experiments, this is not a problem for the
present investigation.
Noise was added using the procedure suggested by Fuchs [28]. Let I
0
be the
expected number of non-attenuated photons traveling along the ray i from (4.1).
Then, the number of attenuated photons I is approximately Poisson distributed
36 Simulation and evaluation
X-ray source Detector element
N
SL
N
SW
N
DH
N
DW
Figure 4.1: In order to realistically simulate the physical acquisition process,
both source and detector are divided into subunits. The final contribution is
calculated as the mean of all sub-contributions.
Table 4.1: Scanning parameters for experiments in following chapters.
Source-Isocenter distance R
S
595mm
Isocenter-Detector distance R
D
490.6mm
Number of channels (Quarter offset) N
ch
672
Number of rows N
rows
48
Number of projections per turn N
proj
1160
Detector subsampling in channel dir. N
DW
3
Detector subsampling in row dir. N
DH
3
Source subsampling along xy-plane N
SW
3
Source subsampling in z-dir N
SL
3
Angle subsampling N
GR
3
Active detector area (xy-plane) d
β
0.8
Active detector area (z-dir) d
z
0.9
Source width (xy) w
S
1.5mm
Source length (z) l
S
12mm
Anode angle η 7

Slicewidth S 1.2mm
Detector height h 57.6mm
Table-feed P 43.2mm
Maximal fan angle β
max
25

Maximal cone angle κ
max
2.77

Pitch angle
a
γ
P
1.04

a
This is the cone angle of a detector covering exactly the Tam-Danielsson
window [110]. It is calculated as γ
P
= tan
−1
(P/(4R
S
)) 180/π.
4.2 Phantoms and error/noise measures 37
Table 4.2: Parameters for the rebinning process.
Number of channels N
t
1344
Number of projections N
θ
1160
Parallel displacement per channel Δ
t
0.386mm
Maximal parallel displacement t
max
260mm
Interpolation in azimuthal rebinning Linear
Interpolation in radial rebinning Nuttall
a
a
See Fig. 11 of [72].
with
E[I] = V ar[I] = I
0
k, (4.3)
where k = exp(−p
i
). When this number is large, the distribution of I is ap-
proximately Gaussian with variance k and standard deviation

k. Thus, a noisy
contribution ˜ p
i
can be calculated as
˜ p
i
= −log

k +

kN

0,
1

I
0

, (4.4)
where N(μ, σ) is a sample from the Gaussian distribution with expectation μ and
standard deviation σ.
Scanning parameters for experiments in following chapters are given by Table
4.1. An illustration of the geometry and corresponding parameters is found in
Chapter 2, Fig. 2.6 and 2.7.
As a first step in all experiments presented in this thesis, input data are re-
binned to semi-parallel geometry. In the rebinning process, new parameters de-
scribing sampling in the new semi-parallel geometry are introduced. Rebinning
parameters are shown in Table 4.2.
4.2 Phantoms and error/noise measures
The mathematical test objects for the studies presented in the following chapters,
are the Turbell Clock phantom [110], and the Forbild Head phantom [59]. The
clock phantom, shown in Fig. 4.2 a), is useful for studying cone artifacts, i.e.,
artifacts that occur due to inexactness of a reconstruction method. It spans the
full field of view, and contains sharp edges orthogonal to the z-direction. The head
phantom is shown in Fig. 4.2 b). Due to its sharp details in the inner ear region,
windmill artifacts [94] and streaks are usually produced.
For each reconstruction result presented in the following chapters, the deviation
from the original phantom has been calculated as the Root Mean Squared Error
(RMSE)
σ
e
=

1
[Ω[
¸
i∈Ω
((f
rec
)
i
−(f
phan
)
i
)
2
. (4.5)
38 Simulation and evaluation
a) Clock phantom b) Head phantom
c) Error/noise mask for the
clock phantom
d) Error/noise mask for the
head phantom
Figure 4.2: Axial images of the clock and the head phantoms. The set Ω
2
is
defined as the white regions in c) and d).
Here, f
phan
is a vector containing sampled values of the original phantom and Ω
is the set of voxels taken into account for the measurement. Depending on how
this set is chosen, different types of deviation are measured. For instance, if high-
contrast edges are included in Ω, σ
e
will be dominated by errors caused by edge
smoothing. For error calculations, we have used two different sets:
• Ω
1
consisting of the complete axial image, and
• Ω
2
consisting only of low contrast regions. For the clock and head phantoms,
the sets Ω
2
consist of the white regions in Fig. 4.2 c) and d) respectively.
Noise in the reconstructed images has been calculated with the formula
σ
n
=

1
[Ω[
¸
i∈Ω
((f
noisy
)
i
−(f
noise−free
)
i
)
2
, (4.6)
where f
noisy
represents a reconstruction from noisy data and f
noise−free
represents
a reconstruction from noise-free data, and the summation is carried out over Ω
2
.
4.3 Spatial resolution measurements 39
a) Axial section
b) Coronal section
Figure 4.3: Thorax phantom reconstruction.
Only simulated input data can deliver perfect ground truth for comparison.
However, due to mismatch between the model used for simulations and the phys-
ical scanning situation, some artifacts prevalent in clinical situations cannot be
captured by simulated data experiments. Therefore, we have included two input
data sets acquired from real medical CT systems in our tests. The first dataset
was acquired from a thorax phantom using a system with 672 detector channels
spanning a field of view of 500mm, and 24 1.2mm detector rows. The second
data set was acquired from a head phantom using a system with 736 detector
channels spanning a field of view of 500mm, and 32 1.2mm detector rows. Both
CT systems were running with a pitch factor (table feed/detector height) of 0.75,
i.e. approximately half of the maximum pitch factor 1.4. Reconstructions of these
phantoms are shown in Fig. 4.3 and 4.4.
Even if the purpose of the iterative methods studied in the following chapters
is to produce good results after few iterations, it is interesting to see what the
methods deliver after many iterations. Therefore, besides error and noise measures
described above, we have calculated the Euclidean norm |f
k+1
− f
k
|. Obviously,
a necessary condition for the iterative method to converge is that this norm tends
to zero as k →∞.
4.3 Spatial resolution measurements
Images reconstructed by analytical and iterative methods, respectively, differ in
many ways, e.g. in sensitivity to noise in input data. Another feature, which often
can be traded for noise is the ability to preserve high frequencies throughout the
reconstruction (high spatial resolution). This section describes how spatial resolu-
40 Simulation and evaluation
a) Axial section
b) Sagittal section
c) Coronal section
Figure 4.4: Head phantom reconstruction.
tion has been measured in the experiments to be found in the following chapters.
We will make a difference between in-plane (xy-plane) spatial resolution, repre-
sented by Modulation Transfer Functions (MTFs), and cross-plane (z-direction)
spatial resolution, represented by Slice Sensitivity Profiles (SSPs).
The most direct and common way of measuring the in-plane MTF is to use a
thin wire test object directed orthogonal to the xy-plane [46], which is scanned
and reconstructed to a very fine grid. Since a wire is an approximation of a Dirac
line, the reconstructed axial image represents the axial Point Spread Functions
(PSFs). By taking the Fourier transform of the PSF, the MTF is obtained. In a
similar fashion, a thin plate orthogonal to the z-axis is often used to measure the
cross-plane SSP.
For iterative methods, MTFs and SSPs normally vary with the reconstruction
volume sampling density. Therefore, by using a very fine grid for resolution mea-
surements as prescribed by the wire and plate methods, the measurements typically
do not correctly describe a clinical situation. To avoid resolution measurements
on sampling densities different from the ones used for normal reconstruction, we
employ the edge-based 3D MTF method by Li et al. [62]. This is an extension of
the edge MTF method originally proposed by Judy [42], and has the advantage
of being able to measure resolution in any location, along any direction, without
changing the sampling density. Below follows a step-by-step description (see Fig.
4.5).
1) Collection of contributing samples. Let f(x) be a reconstruction of the
test object, which is a solid sphere with center x
c
and radius R. Furthermore,
assume that the direction along which we want to measure the resolution is
given by the unit vector d. Now, consider the set S
γ
of sample/distance
4.3 Spatial resolution measurements 41
pairs
(f(x
i
), r
i
(x
i
) = |x
i
−x
c
|) ∈ R R (4.7)
for which (i) x
i
are sample points located within the cone defined by
'd, (x
i
−x
c
)/r
i
(x
i
)` > cos γ (4.8)
and (ii) the distances r
i
(x
i
) satisfy R/2 < r
i
(x
i
) < 3R/2. The set S
γ
contains all information needed to proceed with our MTF measurement.
2) Binning. This step reduces the set S
γ
to a sequence ¦s
j
¦
N
1
of equidistantly
located samples that represent the edge response in the direction d. This
procedure is referred to as binning. Let the locations of these samples be
defined by
˜ r
j
=
R
2
+j
R
N
(4.9)
for j = 0, ..., N. Each element in the sequence of equidistant samples ¦s
j
¦
N
1
is now given by
s
j
=
¸
i,w
ij
>0
f(x
i
)w
ij
¸
i,w
ij
>0
w
ij
, j = 0, ..., N (4.10)
where w
ij
= 1 − [r
i
− ˜ r
j
[/(R/N). Note that how we describe and perform
binning here is slightly different from the original presentation in [62]. While
in the original presentation, w
ij
is a step function of [r
i
−˜ r
j
[, we choose w
ij
to depend linearly on the distance [r
i
− ˜ r
j
[.
3) Differentiation. From the edge response ¦s
j
¦
N
0
, an estimate of the PSF
is obtained by means of differentiation. Specifically, the derivate ¦t
j
¦
N−1
1
is
calculated by means of central differences as
t
j
=
s
j+1
−s
j−1
2R/N
, j = 1, ..., N −1. (4.11)
4) Fourier transformation. To obtain the MTF, the last step is to calculate
the Fourier transform of ¦t
j
¦
N−1
1
.
The sphere center position x
c
and direction vector d determine where and
along which direction the measurement is being made. Two other user configurable
parameters are the so called bin size Δ
˜ r
and the zenith angle γ. The bin size must
be small enough to catch the highest reconstructible frequencies, and the zenith
angle must chosen large enough to ensure that each bin receives sufficiently many
contributions.
In our experiments, we used a radius of R = 20mm (except in Fig. 5.13), a
bin size of Δ
˜ r
= 0.15mm, and a zenith angle of γ = 25

. For measuring in-plane
42 Simulation and evaluation

v
s
0
s
1
s
N
r
s
0
s
N
Edge response
Binning
r
t
1
t
N-1
PSF
FFT
f
MTF
D
i
I
I
e
r
e
n
t
i
a
t
i
o
n
Solid sphere Samples
R
Figure 4.5: Illustration of the edge based 3D MTF method by Li et al. [62].
resolution, this method is very similar to the 2D edge method previously used in
[106]. However, for cross-plane resolution measurements, the proposed method
is more indirect than the previously used multiple plate method. Furthermore,
reconstructed images of solid spheres often contain windmill artifacts, which could
corrupt the measurements. Therefore, we have compared SSP measurements using
the 3D MTF method and the multiple plate method illustrated in Fig. 4.6. Fig.
4.7 shows that there are only very small differences.
It seems appropriate to make the following note on edge based versus sur-
face based methods for examining spatial resolution, the latter being represented
by the above-mentioned thin plate method. In linear reconstruction, it is true
that differentiating an edge response is equivalent to directly measuring the point
response (except for errors introduced by the differentiation). However, for non-
linear methods such as those employing non-linear regularization in Chapter 8,
this is no longer true. Therefore, it is sometimes of interest to examine both edge
responses and surface responses. To this end, we have modified the above method
to measure surface responses by replacing the solid sphere in Fig. 4.5 with a shell
of thickness 0.1mm. Reconstructed images of phantoms for the edge based and
surface based method are shown in Fig. 4.8.
4.3 Spatial resolution measurements 43

Figure 4.6: Multiple thin plate measurement of the SSP of a system. By merging
responses from a number of discs, whose exact positions are known, an oversampled
SSP is obtained.
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
Edge 3DMTF
Surface 3DMTF
Multiple thin plates
Figure 4.7: SSP measurements by the edge and surface based 3D MTF methods,
and the multiple thin plate method.
44 Simulation and evaluation
a) Edge phantom b) Surface phantom
Figure 4.8: Reconstructed axial sections of phantoms for edge based and surface
based resolution measurements. The tests objects have a radius of 2cm, and the
shell phantom in b) has a thickness of 0.1mm. Greyscale window ±50HU.
Chapter 5
Regularized iterative
weighted filtered
backprojection
5.1 Description of the reconstruction algorithm
In this chapter, we investigate the possibility to use the iterative scheme illustrated
in Fig. 5.1 for reducing cone artifacts caused by the non-exact Weighted Filtered
Backprojection (WFBP) reconstruction method [100] (see Sec. 2.2). Much of the
material here has previously been published in [106]. The basic idea is to start with
a non-exact reconstruction f
0
, containing cone artifacts. From this reconstruction,
using a projection operator P, projection data are generated and subtracted from
input data. By adding a reconstruction of these difference projection data to f
0
,
a new improved reconstruction f
1
is obtained.
To simplify our presentation of finite-dimensional linear operators and linear
equation systems, we employ the same matrix/vector notation as in Chapter 3.
Let N be the number of voxels and M be the total number of x-ray attenuation
measurements. Furthermore, let p
in
∈ R
M
denote the input data vector and
let f
0
∈ R
N
denote a vector representing an initial voxel volume. The iterative
method, which we call Regularized Iterative Weighted Filtered Backprojection
(RIWFBP) is illustrated in Fig. 5.1, can now be expressed by the following three
steps.
(1) Rebinning to semi-parallel geometry (p
in
→ p
reb
). This step is equivalent
to the rebinning step of the WFBP method. To preserve radial resolution, the
number of channels is doubled. Therefore, the rebinned and preprocessed input
data p
reb
consist of 2M elements.
(2) Prefiltering (p
reb
→p
mod
). In early experiments we noted that a slight low-
pass filtering of input data can help to suppress overshoots in the xy direction in
46 Regularized iterative weighted filtered backprojection
f
diff
f
k
p
k
+
+
+
Q P
-
- )

p
mod
p
in
Rebinning
PreIiltering
Figure 5.1: Illustration of the RIWFBP method. Given an initial image vector
f
0
= Qp
mod
, a sequence of image vectors is generated by the update formula
f
k+1
= f
k
− αS(Q(Pf
k
− p
in
) + βRf
k
). The matrices Q and P correspond to
WFBP method without rebinning and the forward projection operator modeling
the acquisition process respectively.
the reconstruction results [101]. This was also theoretically motivated by Kunze et
al. [54]. They also showed that a resolution modifying filter such as m(t) described
in Chapter 2 can be used with IFBP methods by performing the convolution as
a prefiltering step. Therefore, to change the resolution properties according to
the filter m(t), and to to suppress overshoots along the z axis, the first step is to
calculate pre-processed input data as
p
mod
= Hp
reb
. (5.1)
Here, the matrix H ∈ R
2M×2M
corresponds to a convolution with m(t) along the
t direction, and a convolution with the filter kernel [γ, (1 − 2γ), γ], γ ∈ [0, 1/3],
along the q direction (see Fig. 2.6 and 2.7 for a description of the geometry). The
amplitude spectrum of this filter is illustrated in Fig. 5.2 for different choices of
γ. To obtain boundary data for the convolution to be performed, extrapolation is
done by copying the lowermost and uppermost detector rows.
Due to the divergent beam geometry in the z direction and the extrapolation
described in the previous paragraph, prefiltering in the z direction corresponds
to a spatially dependent filtering of the reconstruction result, and may therefore
result in new artifacts. In our experiments we used γ = 1/20 which corresponds to
a very moderate smoothing as shown in Fig. 5.2. Clearly, this filter does suppress
high frequencies but only with a factor of 0.8. Nevertheless, as we will see in
Sec. 5.4.1, this cautious choice reduces the overshoot amplitude with a factor of
5. At the same time, the artifact level around sharp z edges gets lower than for
noniterative WFBP.
(3) Iterative application of the WFBP method. This is the main step of the
RIWFBP method. Initially, f
0
is set to Qp
mod
, i.e., the nonexact reconstruction
5.1 Description of the reconstruction algorithm 47
0 0.1 0.2 0.3 0.4 0.5
0
0.2
0.4
0.6
0.8
1
γ = 1/5
γ = 1/10
γ = 1/20
γ = 1/40
γ = 0
Normalized frequency
A
m
p
l
i

c
a
t
i
o
n
Figure 5.2: Amplitude spectrum of the filter [γ, (1 −2γ), γ] along the q direction
in step (1) of the RIWFBP method. The black curve indicate the filter used in
our experiments.
result. The exact update step of the RIWFBP method is given by
f
k+1
= f
k
−αS(Q(Pf
k
−p
mod
) +βRf
k
). (5.2)
In this equation, P ∈ R
2M×N
corresponds to a forward projector. Unless otherwise
noted, we employ the Joseph projector [41] (see Sec. 3.3.4). The matrix Q ∈
R
N×2M
represents the WFBP reconstruction method without the initial rebinning
step. This is the key element which distinguishes the RIWFBP from other iterative
methods. Since this nonexact reconstruction operator is used in each update, a
rapid suppression of cone artifacts is obtained. The matrix R ∈ R
N×N
is a high-
pass filter responsible for the regularization of the method, further described in
the next section. The step length α ∈ R and the low-pass filter S ∈ R
N×N
both
correspond to operations controlling the convergence of the method, and as we will
see below, they do not affect the fixed point as long as the matrix S is invertible.
Throughout this chapter, S correspond to the mild lowpass filter [0.05, 0.9, 0.05]
applied along all three dimensions in the reconstruction domain. This allows for a
slightly higher step length α to be used, which in turn leads to faster suppression
of artifacts.
To find the fixed point of (5.2), we set Δ = I − αS(QP + βR). Iterative
application of 5.2 yields
1
f
k
= Δf
k−1
+αSQp
mod
= Δ
2
f
k−2
+ (I +Δ)αSQp
mod
= Δ
k
f
0
+ (I +Δ+ +Δ
k−1
)αSQp
mod
. (5.3)
1
The described technique for deriving the fixed point was previously used by several authors,
e.g. Zeng and Gullberg [122].
48 Regularized iterative weighted filtered backprojection
By expressing the geometrical sum in closed form, we arrive at the expression
f
k
= Δ
k
f
0
+ (I −Δ
k
)(I −Δ)
−1
αSQp
mod
= Δ
k
f
0
+ (I −Δ
k
)(QP+βR)
−1
Qp
mod
. (5.4)
Thus, given that f
0
linearly depends on input data, f
k
will also depend linearly on
input data. If the spectral radius of Δ is strictly smaller than one, the method
will converge to
f

= (QP+βR)
−1
Qp
mod
. (5.5)
The reconstruction result can be seen as an initial nonexact reconstruction with
Q, followed by a correcting filtering with (QP+βR)
−1
. If QP is invertible, β = 0,
and p
mod
= Pf , then f

= f . However, in practice, p
mod
is not contained in the
range of P. Therefore, numerical studies must be made to investigate how aliasing
and noise affect the reconstructed images.
5.2 The regularization operator R
We start this section by considering the case without regularization, i.e. when
β = 0. Due to the bandlimiting operation in the rampfilter and the interpolating
operations in the forward projector and backprojector, the matrix QP works as
a lowpass filter. Obviously, if the reconstruction grid is finer than the detector
sampling grid, QP will have a nonempty null space consisting of high frequency
structures. Even when the null space of QP is empty, the highest frequencies will
be almost completely suppressed. In turn, this means that (QP)
−1
, if it exists, will
amplify high frequencies very much. Thus, an undesirable enhancement of high
frequencies occurs in images created by (5.5). To avoid this undesired enhancement
of high frequencies and noise, a highpass filter βR is added to QP, resulting in
the matrix QP+ βR. The parameter β controls the amount of smoothing made
by the regularization.
To enable approximately independent control of the frequency responses in the
xy plane and z direction, respectively, we construct the matrix R as
R = (β
xy
R
xy

z
R
z
)/C
N
, (5.6)
where β
xy
and β
z
are parameters used for tuning the in-plane and cross-plane
spatial resolution, respectively. C
N
is a normalization constant mainly introduced
for consistency with material in the following chapters. The matrices R
xy
and R
z
are given by
R
xy
∼ [−1, 2, −1]
x
∗ B
y
∗ B
z
+ [−1, 2, −1]
y
∗ B
x
∗ B
z
(5.7)
and
R
z
∼ [−1, 2, −1]
z
∗ B
x
∗ B
y
. (5.8)
5.3 Local region of interest (LROI) 49
Here, the ∼ means that multiplication with the left-hand side matrices corresponds
to the convolution operations on the right-hand side. B
x
, B
y
, and B
z
denote one-
dimensional lowpass filters in the x, y, and z directions, respectively. These filters
have the form [λ, (1 − 2λ), λ]. To achieve maximal rotation invariance of this
Laplacian filter when β
xy
= β
z
, the parameter λ has been set to λ = 0.093551 as
suggested by Scharr et al. [85]. The normalization constant C
N
has been chosen so
that the center element of the convolution kernel corresponding to (R
xy
+R
z
)/C
N
equals one, i.e., C
N
= 6 (1 −2λ)
2
.
The use of the Laplacian operator as a regularization operator can be motivated
by considering the problem of minimizing the function
z(f ) = g(f ) +
1
2
|∇
s
f |
2
2
(5.9)
where g(f ) is some function measuring the deviation from Pf to input data p, and
1
2
|∇
s
f |
2
2
is a regularization term. The operator ∇
s
∈ R
3N×N
is given by

s
=


T
x

T
y

T
z

T
, (5.10)
i.e., it calculates spatial derivatives in the x, y, and z directions respectively.
Differentiation of z(f ) yields the gradient descent step
f
k+1
= f
k
−α∇z(f ) = f
k
−α(∇g(f
k
) +∇
T
s

s
f
k
). (5.11)
Thus, using the Laplacian ∇
T
s

s
as regularization operator is equivalent to search-
ing for a solution that has low first order derivatives.
In the following chapters, we will also use the equivalent regularization operator
expression
Rf =
N
¸
i=1
N
¸
j=1
d
ij
(f
i
−f
j
)e
i
, (5.12)
where e
i
is the basis vector corresponding to one pixel, and the coefficients d
ij
are
used to describe the neighborhood operation.
5.3 Local region of interest (LROI)
One weakness of iterative reconstruction methods in general is their lacking ability
to reconstruct only a small local region of interest (LROI), partially or completely
contained within the support of the object of interest. This problem is illustrated
in Fig. 5.3. An iterative method uses the difference between projections of recon-
structed images and input data projections to calculate the next estimate. If a
LROI is used, a large mismatch will be observed and erroneously be corrected for.
In contrast to most iterative reconstruction methods, the RIWFBP can recon-
struct a LROI without any preprocessing of input data. However, in this case,
50 Regularized iterative weighted filtered backprojection

v

Iull
()

LROI
()
Figure 5.3: Illustration of the LROI problem. Although the LROI has been
reconstructed correctly, there is a considerable difference between p
full
and p
LROI
.
Thus, an iterative method will observe a difference and erroneously “correct” for
it.
artifacts objects outside the LROI are not suppressed. To understand why LROI
reconstruction is possible, consider partitioning of input data into p
LROI
consist-
ing of contributions from the LROI, and p
LROI
( for complement) consisting of
contributions from outside the ROI, i.e. p = p
LROI
+ p
LROI
. Then, the fixed
point of the RIWFBP iteration is (see (5.5))
f

= (QP+βR)
−1
(Qp
LROI
+Qp
LROI
). (5.13)
If Q would be exact, Qp
LROI
would be zero and f

= (QP + βR)
−1
Qp
LROI
as desired. However, since Q is non-exact, non-local cone artifacts from objects
outside the LROI will distort the result.
A solution to the LROI problem for iterative reconstruction methods in general
has been suggested by Ziegler et al. [127]. The basic idea is to preprocess input
data as shown in Fig. 5.4. Initially, the full Region of Interest (ROI) image is
reconstructed using a simple reconstruction operator Q. From the full ROI image,
the embedded LROI is removed and new projection data are generated using the
projection operator P. The difference between the original and new projection
data is suitable for LROI reconstruction. Since the suggested method removes
most structures outside the LROI, cone artifacts caused from objects outside the
LROI are suppressed by this method.
5.4 Experiments
5.4.1 Artifacts and noise measurements
Fig. 5.5 b) shows an axial image of clock phantom reconstructed with the WFBP
method from data generated as described in Sec. 4.1. In this experiment, regu-
larization and prefiltration parameters were selected so that the spatial resolution
5.4 Experiments 51

-
Q
RIWFBP
Set LROI ÷ 0
P
LROI
Full ROI
Reconstruction
Preprocessing
Figure 5.4: Preprocessing of projection data for LROI reconstruction.
is approximately equal for non-iterative and iterative results. Besides aliasing
artifacts known from two-dimensional reconstruction (see e. g. Kak and Slaney
[45], Chapter 5), the WFBP method seems to generate a mixture of two artifacts:
slowly varying artifacts which we call cone artifacts, and more high-frequent arti-
facts commonly known as spiral or windmill artifacts.
After one iteration (Fig. 5.5 c)), the visual appearance of the reconstructions
is better. Displayed in the 40HU window, the low-frequency cone artifacts are
difficult to spot and the intensity of the windmill artifacts has been reduced. A
similar but less pronounced improvement can be observed between the f
2
and f
3
results. Further iterations result only in a small additional reduction of windmill
artifacts and enhancement of high frequencies.
Fig. 5.6 shows the error measures σ
e1
(whole image) and σ
e2
(without edges)
for the clock phantom experiments. The results observed visually in the previous
paragraph are confirmed. Since the method has been tuned not to affect spatial
resolution properties, there is no notable change in σ
e1
as a function of the iter-
ation number. On the other hand, we observe a reduction of σ
e2
from 6HU to
approximately 3HU in the first iteration. This is followed by smaller reductions in
successive iterations.
52 Regularized iterative weighted filtered backprojection
a) Sampled phantom b) WFBP reconstruction
c) RIWFBP, f
1
d) RIWFBP, f
2
e) RIWFBP, f
3
f) RIWFBP, f
4
Figure 5.5: RIWFBP reconstructions of the clock phantom. Reconstruction
parameters: Q = 0.7, β = 0.52, β
xy
= 1.0, β
z
= 1.5. Greyscale window ±20HU.
5.4 Experiments 53
0 5 10
50
50.1
50.2
50.3
Iteration number
R
M
S
E
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
0 5 10
0
1
2
3
4
5
6
Iteration number
R
M
S
E
,
w
i
t
h
o
u
t
e
d
g
e
s
(
H
U
)
Figure 5.6: RMSE measurements σ
e
on the whole image and on low contrast
regions. Note that σ
e
for the whole image (left plot) stays virtually constant.
Fig. 5.7 shows close-ups of WFBP and RIWFBP (f
5
) reconstructions of the
clock phantom. The RIWFBP reconstructions have been made with and without
the prefiltering described in (5.1). Prefiltering initially reduces the spatial resolu-
tion in the z-direction. However, after five iterations, and having found a suitable
value for the regularization parameter β
z
from (5.6), this blurring seems to be fully
compensated for as shown in Fig. 5.7 c). Comparing Fig. 5.7 b) with 5.7 c), it
seems that this initial reduction of certain high frequencies protects the pre-filtered
f
5
-image from the overshoots in Fig. 5.7 b). Pre-filtering reduces the overshoot
amplitude with more than a factor of five. In previous work, we observed that
overshoots such as the one shown in Fig. 5.7 b) cannot be fully suppressed by
simply increasing β
z
[102].
Fig. 5.8 shows WFBP and RIWFBP(f
5
) reconstructions of a physical thorax
phantom. Due to the low cone angle used in this CT system, there are no cone
artifacts. However, many other types of artifacts exist in the WFBP reconstruc-
tions: streak artifacts, nonlinear partial volume effects, and windmill artifacts.
Similar to the experiments with the clock phantom, we note a slight reduction
of windmill artifacts when comparing the WFBP and RIWFBP reconstructions.
Other artifacts seem to be unaffected by the iterative scheme.
In the thorax phantom experiment, a we employed a resolution modifying filter
enhancing certain low frequencies and suppressing certain high frequencies. An
interesting observation is that, at least from visual inspection of edges Fig. 5.8
c) and d), the frequency response seems to be the same for the WFBP and RI-
WFBP reconstructions. Thus, the resolution modifying filter results in predictable
results not only for WFBP but also for the iterative RIWFBP method. A more
thorough verification of this property is found in the next section, where we study
modulation transfer functions of the reconstruction methods.
Fig. 5.9 shows WFBP and RIWFBP reconstructions of a physical head phan-
tom. Again, windmill artifacts appear in the result. However, since the cone angle
is higher than for the thorax phantom experiment, there are also cone artifacts.
54 Regularized iterative weighted filtered backprojection
a) WFBP reconstruction b) RIWFBP f
5
, no prefilter
c) RIWFBP f
5
, prefilter
Figure 5.7: Zoomed in reconstructions five o’clock sphere close to the edge or-
thogonal to the z-axis. Reconstruction parameters: Q = 0.7, β = 0.52, β
xy
=
1.0, β
z
= 1.5. Greyscale window ±20HU.
These artifacts appear as slowly varying horizontal dark and bright streaks in the
sagittal image in Fig. 5.9 a), and as erroneously dark and bright regions in the
axial image in Fig. 5.9 c). As shown in Fig. 5.9 b) and d), these artifacts are effi-
ciently suppressed by the iterative loop. Also the windmill artifacts are reduced,
but this reduction is more moderate.
Fig. 5.10 shows σ
n
measurements on reconstructions of the clock phantom as
a function of iteration number. Initially, the noise drops from 15.3HU to 14.8HU.
This is mainly because prefiltered input data are not used for the initialization,
but for subsequent iterations. As the iteration number increases, the noise slowly
increases, and after ten iterations, the differences in noise level between WFBP
and RIWFBP are relatively small.
5.4 Experiments 55
a) WFBP reconstruction
b) RIWFBP f
5
c) Zoomed WFBP d) Zoomed RIWFBP f
5
Figure 5.8: Thorax phantom reconstructions. Reconstruction parameters: Q =
0.7, β = 0.52, β
xy
= 1.0, β
z
= 1.5. Greyscale window ±100HU.
56 Regularized iterative weighted filtered backprojection
a) Sagittal image, WFBP
b) Sagittal image, RIWFBP f
5
c) Axial image, WFBP d) Axial image, RIWFBP f
5
Figure 5.9: Physical head phantom reconstructions. Reconstruction parameters:
Q = 0.7, β = 0.47, β
xy
= 1.0, β
z
= 1.4. Greyscale window (40 ± 75)HU. Solid
circles/ellipses indicate cone artifacts. Dashed circles indicate windmill artifacts.
The regularization parameters were selected so that the spatial resolution of the
iterative reconstruction was at least as good as for the WFBP reconstruction.
5.4 Experiments 57
0 1 2 3 4 5 6 7 8 9 10
14.8
14.85
14.9
14.95
15
15.05
15.1
15.15
15.2
15.25
15.3
Iteration number
σ
n
,
(
H
U
)
Figure 5.10: Noise measurements on clock phantom reconstructions.
5.4.2 Spatial resolution
Fig. 5.11 shows MTFs for the WFBP reconstruction, and for RIWFBP with
β = 0.52. In the left figure, β
xy
assumes the values 0.0, 1.0 and 2.0 respectively,
while β
z
is kept constant equal to 1.0. We see that β
xy
= 0.0 gives a slightly
sharper result than the WFBP reconstruction, and that β
xy
= 2.0 gives a slightly
more blurred result than the WFBP reconstruction, while β
xy
= 1.0 gives an MTF
that is very similar to the WFBP MTF. In the right figure, β
xy
is kept constant
while β
z
varies. As expected, this affects the MTF only to a very small extent.
During acquisition and rebinning, data corresponding to rays going through
the peripheral parts of the field of view are blurred more than data corresponding
to rays going through the isocenter. Therefore, it is interesting to see how the
MTFs of WFBP and RIWFBP change as the test object is moved away from the
isocenter. In Fig. 5.12, radial MTFs are shown at the isocenter, and 140mm away
from the isocenter. For a translation of 140mm, the frequency ρ
50%
(WFBP) for
which the MTF of WFBP method has dropped to 50%, is reduced from 3.6HU
to 2.7HU. For amplifications greater than 10%, the relative difference between the
WFBP and RIWFBP f
5
MTF is smaller than 2%, which makes the corresponding
curves overlap each other in Fig. 5.12.
To verify that the RIWFBP in combination with resolution modifying filters
produces expected results in terms of MTFs, we have measured the MTF for one
such filter. The Fourier transform of this filter is shown as the dashed line in
Fig. 5.13. According to the Fourier slice theorem [70], the MTF of WFBP with
a resolution modifying filter should equal the product of the MTF of the WFBP
without a resolution modifying filter m(t) and the Fourier transform of the filter.
58 Regularized iterative weighted filtered backprojection
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
WFBP
RIWFBP, f
5
, β
xy
= 0.0
RIWFBP, f
5
, β
xy
= 1.0
RIWFBP, f
5
, β
xy
= 2.0
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
WFBP
RIWFBP, f
5
, β
z
= 0.0
RIWFBP, f
5
, β
z
= 1.5
RIWFBP, f
5
, β
z
= 3.0
Figure 5.11: MTF measurements for different choices of regularization parame-
ters. In the left image, β
z
is constant equal to one, while β
xy
varies. Note that the
WFBP and RIWFBP β
xy
= 1.0 curves overlap. In the right image, β
xy
is kept
constant equal to one while β
z
varies. For all plots, β is equal to 0.52.
0 1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Frequency (lp/cm)
M
T
F
WFBP, isocenter
RIWFBP, f
5
, isocenter
WFBP, 140mm off center
RIWFBP, f
5
, 140mm off center
Figure 5.12: MTF measurements in the iso-center and 140mm away from the
iso-center. Regularization parameters: β = 0.52, β
xy
= 1.0, β
z
= 1.5. The WFBP
and RIWFBP curves overlap both in the isocenter and off-center cases.
5.4 Experiments 59
0 1 2 3 4 5 6 7 8 9 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
WFBP, Ramlak
Mod. filter
Prediction
WFBP, mod
RIWFBP f
5
, mod
Figure 5.13: MTF measurements for the WFBP and RIWFBP method when
using a resolution modifying filter. Note that the “bump” in the MTF for low
frequencies gives rise to relatively slowly decaying overshoots. Therefore, a slightly
larger solid sphere (radius of 30mm instead of 20mm) was required to capture this
structure of the MTF.
This product, which we may call predicted MTF is shown as the dash-dotted line
in Fig. 5.13. Both the WFBP and RIWFBP MTFs follow the predicted MTF
well, except for very low amplifications, and for low frequencies. The deviation for
low frequencies is caused by the discontinuous derivative of the predicted MTF.
To capture this structure, the measurement must be made over a large region. In
this particular case, we used a test object with radius 30mm instead of 20mm.
Fig. 5.14 shows SSPs for WFBP and RIWFBP reconstructions. In the left
column, β
xy
is varied while β
z
is being held constant. As expected, since β
z
is
held constant, the SSPs do not change much in this case. In the right column,
β
xy
is held constant equal to 1.0 while β
z
assumes the values 0.0, 1.5 and 3.0
respectively. Unfortunately, although the prefiltering helps to suppress overshoots,
none of the SSPs of the RIWFBP method resemble that of the WFBP method
very well. Initially, the RIWFBP SSPs drop faster, resulting in a smaller Full
Width at Half Maximum (FWHM). However, in the region where the SSPs have
dropped to less than 1% of the maximal values, the RIWFBP SSPs are wider than
the WFBP SSPs. Changing β
z
mainly affects the amplitude of the tails, rather
than the width.
When choosing a β
z
, we must take into account both the FWHM and the SSPs
tail widths. Considering only FWHM, β
z
= 3.0 gives the best correspondence
between the WFBP SSPs and the RIWFBP SSPs. However, since no regards
have been paid to the wider tails of the RIWFBP SSPs, this choice is unfair in
60 Regularized iterative weighted filtered backprojection
0 0.5 1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
WFBP
RIWFBP, f
5
, β
xy
= 0.0
RIWFBP, f
5
, β
xy
= 1.0
RIWFBP, f
5
, β
xy
= 2.0
0 0.5 1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
WFBP
RIWFBP, f
5
, β
z
= 0.0
RIWFBP, f
5
, β
z
= 1.5
RIWFBP, f
5
, β
z
= 3.0
1 1.5 2 2.5
−0.05
0
0.05
0.1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
1 1.5 2 2.5
−0.05
0
0.05
0.1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
Figure 5.14: SSP measurements for different choices of regularization parame-
ters. In the left column, β
z
is constant equal to one, while β
xy
varies. In the right
column, β
xy
is kept constant equal to one while β
z
varies. For all plots, β is equal
to 0.52.
favor of the RIWFBP. All things considered, we have chosen the regularization
parameter value β
z
= 1.5 for the comparative studies in the previous section.
5.4.3 Higher number of iterations. Long objects.
In Fig. 5.15, the function log
10
(|f
k+1
− f
k
|/C) is plotted against the iteration
number for head phantom reconstructions. This has been done both for the non-
regularized and regularized IWFBP, and over two different images: one located
centrally at z = 0mm, and one located peripherally at z = 56mm (the most
peripheral image properly reconstructed by the WFBP method). In the nonreg-
ularized case, i.e. β = 0, the difference norm decays slowly for both the central
and the peripheral image. After 100 iterations, the norm has decayed to 4 10
−3
of its initial value. In contrast, with regularization, this function drops down to
1.7 10
−4
within 30 iterations for the central image. For the peripheral image,
the decay is slower, and more than 100 iterations seem to be necessary to reach
5.4 Experiments 61
0 20 40 60 80 100
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
Iteration number
l
o
g
1
0
(
|
f
k
+
1

f
k
|
/
C
)
Central image, β = 0.52
Central image, β = 0
Peripheral image, β = 0.52
Peripheral image, β = 0
Figure 5.15: The difference norm log
10
(|f
k+1
− f
k
|/C) plotted against the
iteration number for one central image (z = 0mm) and one peripheral image
(z = 56mm). For each curve, C was selected as max
k
(|f
k+1
−f
k
|).
convergence. Note that this is not a practical problem, since acceptable results
are produced within 5 iterations.
Fig. 5.16 shows the error measures σ
e1
(whole image) and σ
e2
(low contrast
regions) for the “central image” studied in Fig. 5.15, with and without regulariza-
tion. These measurements are consistent with those in Fig. 5.15 in the sense that
with regularization, the measurements practically remain constant after first ten
iterations, while without regularization, noticeable changes occur even after 100
iterations. Starting with σ
e1
, we note that without regularization, this measure
initially drops from 102HU to 95HU during the first ten iterations. In the following
90 iterations, it slowly increases with a total of approximately 0.5HU. With reg-
ularization, σ
e1
remains practically constant except for a small initial reduction.
Now turning our attention to the σ
e2
measure, we see that without regularization,
there is an initial drop followed by a steady increase of the error. With regulariza-
tion, as a sign of reduction of cone artifacts and windmill artifacts, the σ
e2
measure
drops from 10.3HU to 8.7HU in the first ten iterations, and remains constant on
this level.
Fig. 5.17 shows σ
e1
and σ
e2
for the “peripheral image” in Fig. 5.15. These
plots are very similar to those in Fig. 5.17, except for the initial “bump” observable
in all curves except σ
e1
with β = 0. This seem to be an error caused by an initial
overcompensation for differences between Pf and p
reb
near the ends of available
input data. Although the initial increase in σ
e2
error is large in relation to the
existing errors, it is small in relation to the linear attenuation values in the image,
and hardly visible in a 40HU window.
62 Regularized iterative weighted filtered backprojection
0 50 100
94
96
98
100
102
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
1
)
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
0 50 100
8.5
9
9.5
10
10.5
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
2
)
,
l
o
w
c
o
n
t
r
a
s
t
r
e
g
i
o
n
s
(
H
U
)
Figure 5.16: RMSE (σ
e1
and σ
e2
) for head phantom reconstructions at a central
image (z = 0mm).
0 50 100
69
70
71
72
73
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
1
)
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
0 50 100
0
0.5
1
1.5
2
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
2
)
,
l
o
w
c
o
n
t
r
a
s
t
r
e
g
i
o
n
s
(
H
U
)
Figure 5.17: RMSE (σ
e1
and σ
e2
) for head phantom reconstructions at a periph-
eral image properly reconstructible with WFBP(z = 56mm).
Fig. 5.18 shows σ
e1
and σ
e2
for an image that cannot be reconstructed with
the WFBP because of missing data. In this case, both error values are clearly
reduced by the RIWFBP method.
Fig. 5.19 shows sagittal images of Head phantom reconstructions for different
number of iterations, with and without regularization. The dashed lines indicate
the points where the WFBP reconstructions lose intensity due to insufficient data.
Focusing on the central regions of the phantom, we see that most of the improve-
ment take place during the first five iterations. At this point, the regularized results
very much look like smoothed versions of the non-regularized results. However,
when the number of iterations increases further, we notice that the nonregularized
results continue to change, while the regularized remain practically unchanged as
the number of iterations goes from five to fifty.
In the peripheral regions where the long object problem might show up, the
behavior of the RIWFBP and the WFBP methods are quite different. Except
from the initial slight increase of artifacts in images close to regions that are not
5.4 Experiments 63
0 50 100
65
70
75
80
85
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
1
)
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
0 50 100
0
5
10
15
20
25
β = 0.52
β = 0
Iteration number
R
M
S
E
(
σ
e
2
)
,
l
o
w
c
o
n
t
r
a
s
t
r
e
g
i
o
n
s
(
H
U
)
Figure 5.18: RMSE (σ
e1
and σ
e2
) for head phantom reconstructions at a periph-
eral image not properly reconstructible with WFBP(z = 59mm).
reconstructible with the WFBP method, there seems to be a steady increase in
reconstructible region during the first fifty iterations. This effect is even more
visible in Fig. 5.20, which shows axial images at z = 60mm. In this figure, we see
that after five iterations, an improved result is obtained, where the dark region
has been overcompensated for, and turned into a slightly too bright region. As
the number of iterations further increase, this defect disappears.
A related study on iterative application of the PI-method [11, 110] was made
by Magnusson et al. [65]. Those results are consistent with the results here except
for the initial increase of errors that we observed. One reason for this inconsistency
could be that the WFBP method, in contrast to the PI-method, incorporates data
outside the Tam-Danielsson window.
5.4.4 Local region of interest
The LROI reconstruction method was described in Sec. 5.3. For the initial full
reconstruction over ROI we are free to choose any reasonable non-exact algorithm.
In our case it was reasonable to use WFBP.
To examine the possibilities to reduce artifacts caused by objects located both
inside and outside the LROI, we have selected a cylindrical LROI just above the
outer nine o’clock sphere in the clock phantom. Fig. 5.21 a) shows a full ROI
WFBP reconstruction zoomed in over this LROI. The image is distorted by arti-
facts caused by one solid sphere located in the LROI, and one solid sphere located
just below the LROI. After five iterations, the cone artifacts are suppressed and
the windmill artifacts are slightly alleviated as shown in Fig. 5.21 b).
Fig. 5.21 c) and d) show WFBP and RIWFBP LROI reconstructions without
the preprocessing of input data described in Fig. 5.4. Obviously, the WFBP
LROI reconstruction is identical to the WFBP full ROI reconstruction. After
five iterations, the cone artifact caused by the solid sphere inside the LROI is
suppressed. However, as expected, the other artifact remains unaffected.
64 Regularized iterative weighted filtered backprojection
a) Phantom b) WFBP reconstruction
c) RIWFBP, f
5
, β = 0 d) RIWFBP, f
5
, β = 0.52
e) RIWFBP, f
10
, β = 0 f) RIWFBP, f
10
, β = 0.52
g) RIWFBP, f
50
, β = 0 h) RIWFBP, f
50
, β = 0.52
Figure 5.19: Sagittal images of RIWFBP reconstructions of the head phantom.
Greyscale window ±50HU.
5.5 Discussions and conclusions 65
a) WFBP reconstruction b) RIWFBP, f
5
c) RIWFBP, f
10
d) RIWFBP, f
50
Figure 5.20: Axial images of RIWFBP reconstructions of the head phantom at
z = 60mm. Greyscale window ±50HU.
Fig. 5.21 e) and f) show WFBP and RIWFBP LROI reconstructions on LROI
preprocessed input data. Clearly, the pre-processing suppresses artifacts non-local
artifacts caused by objects located outside the LROI. After five iterations, the
other cone artifact has also been suppressed, and the result look very similar to
the full ROI result.
5.5 Discussions and conclusions
The main advantage of the proposed RIWFBP method is its high rate of conver-
gence relative to other iterative methods. For full resolution medical CT input
data at cone angles up to ±2.78

, cone artifacts are perfectly suppressed within
three iterations, and there is also a clear alleviation of windmill artifacts. For
comparison, De Man et al. [18] state that for convergence, current state of the
66 Regularized iterative weighted filtered backprojection
a) Full ROI, f
0
b) Full ROI, f
5
c) LROI, without preproc., f
0
d) LROI, without preproc., f
5
e) LROI, with preproc., f
0
f) LROI, with preproc., f
5
Figure 5.21: Full ROI and LROI reconstructions of the clock phantom. Greyscale
window ±20HU.
5.5 Discussions and conclusions 67
art methods for statistical reconstruction (iterative coordinate descent (ICD) [109]
and ordered subsets convex (OSC) [52]) require a computation time equivalent to
60 RIWFBP iterations.
Statistical reconstruction methods studied e.g. by Thibault et al. [109] include
weighting factors, which take into account the Poisson distribution of measured
input data. Hereby, the signal to noise ratio can be improved. However, the
weighting makes the corresponding optimization problem more difficult to solve
within a short period of time. The Poisson model may be possible to adopt in
RIWFBP, but it is far from obvious how such weighting can be incorporated into
the RIWFBP framework without increasing the number of iterations considerably.
Recently, Wang et al. [114] presented exact reconstruction methods based on
work by Katsevich [48] and Zou and Pan[128] as potential alternatives to con-
temporary methods in commercial systems. Since exact methods directly produce
results free from cone artifacts, one might think that there would be no need
for non-statistical iterative methods such as the RIWFBP method. However, we
believe that RIWFBP has at least three important advantages over currently pre-
sented exact methods.
• First, existing exact methods are limited to utilizing data from the Tam
window [108]/PI detector [10] or the n-PI detectors [80] (see Fig. 5.22). As
shown in Fig. 5.22, for a pitch factor of 0.75, the dose utilization for an
exact method is less than 50% for the Tam window. Zamyatin et al. [120]
experimentally verified that this affects the signal to noise ratio negatively.
For some approximate methods (Flohr et al. [26], Schechter et al. [88], and
Zamyatin et al. [120]), it has been demonstrated that incorporation of data
outside the Tam window can improve the signal to noise ratio. Also, since
the principles of incorporation of redundant data are similar for the WFBP
and the Adaptive Multiple Plane Reconstruction (AMPR) method (Flohr et
al. [26]), we believe that this is true also for the WFBP method.
• Second, the iterative nature of RIWFBP makes it possible to incorporate
more advanced forward projection models in order to compensate for blurring
effects caused by the finite size of focal spot and detector elements, and the
gantry rotation during acquisition (see the final paragraph).
• Third, RIWFBP does not suffer from the sensitivity in implementations of
derivation, masking, and Hilbert transformation characteristic for the pro-
posed exact methods.
A number of other approaches exist for alleviating wind-mill artifacts in helical
cone-beam CT. Software approaches include interpolation between conjugate rays
during backprojection [38], non-linear filtering [35], and modified linear interpo-
lation [119]. Compared to the RIWFBP method, these methods are faster, since
they are not based on iterative application of projection and backprojection op-
erators. On the other hand, the RIWFBP method has the advantage of not only
68 Regularized iterative weighted filtered backprojection
x
y
z
Figure 5.22: Illustration of the relation between a clinical CT detector, the PI
detector and the 3-PI detector for a pitch factor of 0.75. White: 3-PI detector.
Light gray: clinical CT detector. Dark gray: PI-detector.
alleviating wind-mill artifacts, but also suppressing cone artifacts, making it more
suitable for systems with higher cone angles.
It should be mentioned that besides the above mentioned software approaches
for wind-mill artifact reduction, there also exists a hardware technique already
employed in clinical CT systems [27]. By rapidly moving the focal spot between
two different z-positions, and increasing the number of readouts per turn with
a factor of two, a virtual detector with twice as many rows is created. Hence,
the aliasing during acquisition is reduced, resulting in a considerable alleviation
of wind-mill artifacts. Unfortunately, the moving spot mechanism comes with
increased complexity and costs in many parts of the acquisition system. Therefore,
software solutions for alleviating wind-mill artifacts are still an interesting topic
of investigation.
Our investigations employed a 3 3 3 approximation of a Laplacian for
regularization. There is no strong theoretical motivation for this choice, and there
are many other possibilities, for instance more accurate Laplacian approximations,
and other types of high-pass filters, all of which could be used to further equalize
the frequency response between WFBP and RIWFBP.
An important issue in iterative reconstruction is linear versus non-linear regu-
larization. In this chapter, we have limited our investigations to linear regulariza-
tion because it simplified the convergence analysis. Furthermore, because of the
full linearity of each and every part of the reconstruction system, spatial resolution
measures such as MTF and SSP can be used for analysis and characterization of
the reconstruction method. Another benefit from the linearity is the possibility to
5.5 Discussions and conclusions 69
use resolution modifying filters frequently used in clinical CT systems. Nonethe-
less, non-linear regularization has shown to be useful for preserving high-contrast
resolution while reducing noise at low contrasts (see e. g. Thibault et al. [109]),
and will therefore be the topic of Chapter 8.
As shown in Section 5.4.2, there are spatial variations in the MTF of the
RIWFBP method. These variations follow the variations of the WFBP MTF.
Although not investigated here, similar variations are expected for noise and SSPs.
In contrast to column-action methods such as the ICD [109] method or row-
action methods such as the algebraic reconstruction technique (ART) [31], the
update step in RIWFBP is completely simultaneous. In other words, within each
iteration, there are no dependencies between updates of different voxels or rays.
Therefore, the resulting method is highly parallelizable, and can be efficiently im-
plemented in modern hardware. Furthermore, the problem of finding efficient re-
laxation factors is less critical, and the problem of finding the best projection/voxel
access scheme is non-existent.
During the input data acquisition, measured data are blurred by the finite size
of the focus spot, the finite size of the detector elements, and the gantry rota-
tion during the measurement. Few, if any contemporary reconstruction methods
include compensation of these effects, which results in reconstructed images with
spatial resolutions below the ones to be expected from actual numbers of views and
detector element densities. In iterative reconstruction, it is far more easy to incor-
porate accurate acquisition models. Improving the acquisition model implemented
in the forward operator P is the topic of Chapter 9.
70 Regularized iterative weighted filtered backprojection
Chapter 6
Ordered Subsets IWFBP
(OS-IWFBP)
6.1 Introduction
This chapter is an adapted and reduced version of [102, Chapter 6], presenting
the OS-IWFBP method, which is a non-regularized and accelerated version of the
RIWFBP method. Since the original publication, our investigations have focused
on other topics, mainly because the OS-IWFBP method seemed to be useful only
for extremely high cone angles (κ
max
> 4.8

), which are unlikely to be used for
medical helical cone-beam CT. However, many recent investigations indicate a
need for iterative methods with a higher rate of convergence than the RIWFBP
from the previous chapter. In those cases, the Ordered Subsets (OS) technique
could be an interesting candidate for improving the rate of convergence.
In the following sections, we first describe the OS technique and then move on
to how it is applied in the OS-IWFBP method. This is followed by an experiment
showing how OS-IWFBP efficiently suppresses cone artifacts for high cone angles.
Finally, we summarize the conclusions and discuss possible future investigations.
6.2 The Ordered Subsets (OS) technique
The OS technique for emission tomography was introduced in 1994 by Hudson and
Larkin [39]. They showed that this technique could improve the rate of convergence
for the emission tomography MLEM method [90] with at least a factor 16 without
any appreciable loss of image quality. Later, the ordered subsets technique was
applied to algorithms for transmission tomography by Kamphuis and Beekman
[47] and Erdo˘ gan and Fessler [24].
Mathematical descriptions and proofs of convergence for sequential methods
such as ART, SART, and OS methods were given by Jiang and Wang in [40]. Here,
72 Ordered Subsets IWFBP (OS-IWFBP)
Simultaneous mode:
OS mode:

in

1
.

2
. ... .

12

in ,1

1
.

4
.

7
.

10

in ,2

2
.

5
.

8
.

11

1

2

3

4

5

6

7

8

9

10

11

12

in,3

3
.

6
.

9
.

12

Figure 6.1: Simple illustration of OS in the case of 12 projection angles and
three subsets. Each arrow represents a parallel projection angle θ. A simultaneous
method would use all projection data p
in
in each update. In the illustrated OS
case, the full update would be split into three partial updates working on p
in,1
,
p
in,2
, and p
in,3
respectively.
we give a brief presentation of the OS technique for least squares minimization (see
Section 3.1), followed by a presentation of the OS-IFBP method.
In an OS method, the input data p
in
are partitioned into equidistantly sampled
subsets p
in,l
, l = 1, ..., L with respect to the rotation angle θ (see Fig. 6.1). For each
subset, a projection matrix P
l
is defined. This matrix only generates projection
data for the subset l, i.e, P
l
consists only of the rows of P corresponding the subset
l. With an index sequence ¦i
k
¦

k=1
specifying the order in which the subsets shall
be applied, the update step becomes
f
k+1
= f
k

k
P
T
i
k
(p
in,i
k
−P
i
k
f
k
). (6.1)
For the case of disjoint subsets, Jiang et al. [40] showed that if the index
sequence is periodic, i.e. i
k+nL
= i
k
, n ∈ N, and the step lengths α
k
satisfy the
conditions
lim
k→∞
α
k
= 0, and

¸
k=0
α
k
= +∞, (6.2)
then the method defined by (6.1) converges towards a minimum norm minimizer
of |Pf −p
in
|
2
plus a projection of the initial image f
0
on the null space N(P) =
R(P
T
)

.
Although the above result holds independently of subset ordering, the ordering
affects the rate of convergence. Previous investigations have shown that while ob-
viously inefficient orderings such as linearly increasing θ-offset (¦i
k
¦ = ¦1, 2, 3, ...¦)
should be avoided, there are very small differences between common ordering
schemes, such as random ordering, the golden ratio based ordering [49], and the
ordering scheme by Beekman and Kamphuis [2] which we will use in the following
sections.
6.3 OS-IWFBP 73
6.3 OS-IWFBP
Similarly to the RIWFBP, the first step in the OS-IWFBP method is rebinning of
the input data p
in
to semi-parallel data p
reb
. These data are then divided into L
subsets p
reb,1
, ..., p
reb,L
. For each subset, a specific projector/reconstructor pair
P
l
/Q
l
is created. The iteration of OS-IWFBP is given by
f
k+1
= f
k

k
Q
i
k
(p
reb,i
k
−P
i
k
f
k
). (6.3)
Assuming that the ordering ¦i
k
¦ is periodic with period L, and that all subsets
are applied during one period, we define one full iteration as the collection of L
consecutive sub-iterations.
High values of α (close to one) imply efficient suppression of cone artifacts.
However, as we will see in the next section, unless very low step lengths α
k
are
used (α
k
= 0.15 for ten subsets), divergent high frequency structures develop
during the iterations. To allow for higher step lengths, row-wise low-pass filtering
of projection data differences (p
reb
−P
i
k
f
k
) is incorporated in the reconstructions
Q
l
. The kernel used for low-pass filtering is a truncated and sampled Gaussian
given by
K
σ
[m] =

1
C
exp


m
2
σ
2

, m = −
M−1
2
, ...,
M−1
2
0 , otherwise
, (6.4)
where C =
¸
m
exp


m
2
σ
2

so that the DC-component of the filter equals one.
We use the ordering scheme suggested by Beekman and Kamphuis [2]. This
ordering scheme is illustrated in Fig. 6.2. Given a number of subset indices
i
1
, i
2
, ..., i
n
, the next index i
n+1
is chosen as close as possible to the center of the
largest gap of unused subsets. If several gaps of the same size exist, the center
of the gap located as far away as possible, i.e. as close to LΔ
θ
/2 as possible,
is chosen. Sometimes several subsets satisfy the last condition (see for instance
Fig. 6.2 e)). Then the choice between the remaining centrally located subsets is
random.
6.4 Experiments
To examine how reconstruction errors depend on the number of subsets L, and
step length α
k
= α, we have performed two-dimensional experiments with the
Forbild thorax phantom [97] shown in Fig. 6.3. The underlying assumption is that
unwanted artifacts generated by the OS technique will prevail in the same way in
the three-dimensional case. Scanning geometry parameters for all experiments in
this chapter are listed in Table 6.1.
Ten full iterations per configuration have been performed. Visual inspection
reveals that for high values of α, certain noise-like artifacts appear in the OS-
IWFBP results (see Fig. 6.6 c)). As a numerical measure of these artifacts, the
74 Ordered Subsets IWFBP (OS-IWFBP)

1
2
3
4
5
6
0 0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6
0
1
2
3
4
5
6


a) i
0
= 0 ⇒largest gap = ¦1...6¦ ⇒i
1
= 4
b) i
1
= 4 ⇒largest gap = ¦1...3¦ ⇒i
2
= 2
c) i
2
= 2 ⇒largest gap = ¦5...6¦ ⇒i
3
= 6
d) i
3
= 6 ⇒equally sized gaps = ¦1¦, ¦3¦, ¦5¦ ⇒i
4
= 3
e) i
4
= 3 ⇒equally sized gaps = ¦1¦, ¦5¦ ⇒i
5
= 1 or i
5
= 5
Figure 6.2: Illustration of the ordering scheme suggested by Beekman and Kam-
phuis [2]. Given a number of subset indices i
1
, i
2
, ..., i
n
, the largest gaps of unused
subsets are identified. The new index i
n+1
is chosen as close as possible to the gap
center located as far away from the subset i
n
as possible.
Figure 6.3: Axial image of the Forbild thorax phantom [97] at z = 0mm.
6.4 Experiments 75
Table 6.1: Scanning parameters for experiments on OS-IWFBP.
Source-Isocenter distance R
S
570mm
Number of channels (Quarter offset) N
ch
336
Number of rows N
rows
1/64/128
Number of projections per turn N
proj
580
Detector height h 1.5mm, 96mm, 192mm
Table-feed P 0mm, 96mm, 192mm
Maximal cone angle κ
max
0

/4.8

, 9.6

error (repetition of (4.5))
σ
e
=

1
[Ω[
¸
i∈Ω
((f
rec
)
i
−(f
phan
)
i
)
2
(6.5)
has been calculated over low contrast regions.
Fig. 6.4 shows σ
e
measurements for different number of subsets and step
lengths without the lowpass filter in (6.4). For comparison, the black solid line with
ring markers show the corresponding error for RIWFBP with β = 0. Clearly, the
highest allowable α varies strongly with the number of subsets. For the different
subset configurations included in this experiment, α = 1/L seem to result in
approximately the same error curve as the RIWFBP method with β = 0.
To examine the effect of the lowpass filter in (6.4), experiments have been
performed with five and ten subsets. The σ-values determining the amount of
low-pass filtering were set to 1, 2, and 4. Error measurements σ
e
with respect to
the number of iterations are shown in Fig. 6.5. In all four cases, the amounts of
artifacts at low contrast regions are reduced as σ is increased. For five subsets
with α = 0.50 and ten subsets with α = 0.25, the previously divergent sequences
are turned into more stable sequences showing no signs of divergence during the
first ten iterations.
When σ is further increased, the amount of artifacts at low contrast regions
drops below those introduced by RIWFBP with β = 0. This is mainly because
increasing σ also increases the required number of iterations for high frequencies
to be reconstructed. Therefore, for large σ values we expect the σ
e
values to slowly
increase to levels similar to those of the RIWFBP with β = 0.
The RMSE gives no information about the structure of the errors. Therefore,
to see what the numbers presented in Fig. 6.4 and 6.5 mean in terms of visual
artifacts, axial images of reconstructions with ten subsets and two full iterations are
shown in Fig. 6.6. These images show that for α = 0.15, the OS-IWFBP produces
similar images to those of the RIWFBP without regularization. However, when
α is increased from 0.15 to 0.25, strong ringing artifacts appear in the result. By
using the lowpass filter from (6.4) with σ = 1, these artifacts are clearly suppressed.
To study cone artifact suppression, two systems were considered: one 64-row
system with a maximum cone angle κ
max
= 4.8

and one 128-row system with
76 Ordered Subsets IWFBP (OS-IWFBP)
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.0
OS−IWFBP, α÷1.0
OS−IWFBP, α÷0.7
OS−IWFBP, α÷0.5
OS−IWFBP, α÷0.2
a) 2 subsets
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.50
OS−IWFBP, α÷0.35
OS−IWFBP, α÷0.20
OS−IWFBP, α÷0.10
b) 5 subsets
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.25
OS−IWFBP, α÷0.15
OS−IWFBP, α÷0.10
OS−IWFBP, α÷0.05
c) 10 subsets
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.10
OS−IWFBP, α÷0.07
OS−IWFBP, α÷0.05
OS−IWFBP, α÷0.02
d) 29 subsets
Figure 6.4: RMSE/σ
e
measurements on low contrast for RIWFBP with β = 0,
and OS-IWFBP with different number of subsets and step lengths.
6.4 Experiments 77
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.35,NoIilter
OS−IWFBP, α÷0.35, σ÷1
OS−IWFBP, α÷0.35, σ÷2
OS−IWFBP, α÷0.35, σ÷4
a) 5 subsets, α = 0.35
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.5,NoIilter
OS−IWFBP, α÷0.5, σ÷1
OS−IWFBP, α÷0.5, σ÷2
OS−IWFBP, α÷0.5, σ÷4
b) 5 subsets, α = 0.50
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.15,NoIilter
OS−IWFBP, α÷0.15, σ÷1
OS−IWFBP, α÷0.15, σ÷2
OS−IWFBP, α÷0.15, σ÷4
c) 10 subsets, α = 0.15
0 2 4 6 8 10
2
2.5
3
3.5
4
Number of full iterations
R
M
S
E
(
H
U
)
RIWFBP, α÷1.00
OS−IWFBP, α÷0.25,NoIilter
OS−IWFBP, α÷0.25, σ÷1
OS−IWFBP, α÷0.25, σ÷2
OS−IWFBP, α÷0.25, σ÷4
d) 10 subsets, α = 0.25
Figure 6.5: RMSE/σ
e
measurements on low contrast regions for different degrees
of low-pass filtering.
78 Ordered Subsets IWFBP (OS-IWFBP)
a) RIWFBP, α = 1.0, β = 0 b) OS-IWFBP, α = 0.15, No LP
c) OS-IWFBP, α = 0.25, No LP d) OS-IWFBP, α = 0.25, σ = 1
Figure 6.6: Blown up two-dimensional reconstructions of the Forbild Thorax
phantom [97], showing high-frequency “ringing” artifacts, and how they can be
suppressed by the low-pass filter in (6.4). The number of subsets is ten. Greyscale
window ±50HU.
6.5 Discussions and conclusions 79
a maximum cone angle κ
max
= 9.6

. Fig. 6.7 shows axial reconstructions of the
clock phantom for κ
max
= 4.8

. In this case, the OS technique does not seem to
offer much improvement to the result obtained with RIWFBP. Already after one
iteration, decent suppression of cone artifacts is obtained without OS. The OS-
IWFBP result is only marginally better. However, for the higher cone angle ±9.6

shown in Fig. 6.8, the difference between the methods is more pronounced. After
two iterations, there are clearly visible cone artifacts remaining in the RIWFBP
result, while the image produced by OS-IWFBP is close to perfect in this respect.
6.5 Discussions and conclusions
The main motivation for combining the OS technique with the RIWFBP algo-
rithm is to reduce the required number of iterations to suppress cone artifacts, or
in other ways improve the image quality. Indeed, for κ
max
= 9.6

the OS-IWFBP
clearly performs better than the RIWFBP method in terms of cone artifact reduc-
tion. However, for cone angles smaller 4.8

, the differences in efficiency are less
pronounced. Therefore, it is difficult to motivate the use of OS for improved cone
artifacts suppression for low cone angles.
To obtain maximal suppression of cone artifacts, the step length α should be
close to one. However, the maximum allowable step length seems to be inversely
proportional to the number of subsets. One way to allow for higher step lengths
and thereby improve the efficiency is to lowpass filter projection data in the update
step according to (6.4). Our results show that by applying a Gaussian lowpass
filter with σ = 1 in the update step, the maximum allowable α can be increased
from 0.35 to 0.50 for five subsets, and from 0.15 to 0.25 for ten subsets.
Although regularization has not yet been investigated for the OS-IWFBP, we
note that it can be added in the same way as for the RIWFBP, resulting in the
update step
f
k+1
= f
k

k
Q
i
k
(p
in,i
k
−P
i
k
f
k
−βRf
k
). (6.6)
This is motivated by recognizing that Q
l
≈ Q. Thus, for a low number of sub-
sets L, the update step (6.6) is approximately equal to the RIWFBP update step.
However, as the number of subsets grows, the approximation becomes more inac-
curate, and multiplication with R becomes more expensive relative multiplication
with P
l
and Q
l
. For example, if the computational cost of R is one tenth of the
cost of P
l
and Q
l
, the cost of a full iteration increases with 50% for ten subsets
and 100% for twenty subsets. Obviously, this increase in computational cost per
iteration needs to be taken into account when examining improvements in rate of
convergence.
As shown in the previous section, the OS-IWFBP efficiently improves the recon-
struction results during the first iterations. However, since there is no mechanism
to control the step lengths of the method, OS-IWFBP is not convergent. Instead,
80 Ordered Subsets IWFBP (OS-IWFBP)
(a) RIWFBP, β = 0, f
0
(b) OS-IWFBP, f
0
(c) RIWFBP, β = 0, f
1
(d) OS-IWFBP, f
10
(e) RIWFBP, β = 0, f
2
(f) OS-IWFBP, f
20
Figure 6.7: Clock phantom reconstructions with RIWFBP (β = 0) and OS-
IWFBP with 10 subsets. Cone angle κ ≈ ±4.8

. For the OS-IWFBP recon-
structions, α = 0.25 and σ = 1. Greyscale window ±50 HU. Note that the
computational cost is approximately the same for c) and d), and for e) and f).
6.5 Discussions and conclusions 81
(a) RIWFBP, β = 0, f
0
(b) OS-IWFBP, f
0
(c) RIWFBP, β = 0, f
1
(d) OS-IWFBP, f
10
(e) RIWFBP, β = 0, f
2
(f) OS-IWFBP, f
20
Figure 6.8: Clock phantom reconstructions with RIWFBP (β = 0) and OS-
IWFBP with 10 subsets. Cone angle κ ≈ ±9.6

. Greyscale window ±50 HU. Note
that the computational cost is approximately the same for c) and d), and for e)
and f).
82 Ordered Subsets IWFBP (OS-IWFBP)
it either diverges, or enters a limit cycle, i.e., an infinite cycle of recurring recon-
struction results. One way to avoid these phenomenons is to use the OS-IWFBP
for a couple of initial iterations, and then either to switch to another method that
guarantees convergence [98], or to successively reduce the number of subsets [2].
In Chapter 8 on nonlinear regularization and Chapter 9 on improved data
acquisition modeling, it is shown that these techniques can improve images in
terms of spatial resolution versus noise. However, this comes to the cost of an
increased number of required iterations. Therefore, an interesting topic for future
research is to examine if OS can be used to improve the rate of convergence in
these cases.
Chapter 7
Relation between RIWFBP
and least squares methods
7.1 Motivation
The RIWFBP method presented in Chapter 5 is intuitively appealing: since Q
is an approximate inverse of P, already f
0
= Qp
reb
corresponds to a relatively
good reconstruction result. Thus, the RIWFBP method is a practical method for
improving reconstructed images within very few iterations. Still, from a theoretical
point of view, the method is less satisfying. In particular, it is difficult to relate
and compare RIWFBP to other iterative methods. To this end, we suggest a
new method called WLS-IFBP, for which such comparison is possible. This new
method is compared to the RIWFBP method with respect to artifact reduction,
noise, and spatial resolution.
If there is a fixed point of the RIWFBP iteration, it can be expressed as
(repetition of (5.5))
f

= (QP+βR)
−1
Qp
reb
(7.1)
which trivially minimizes the quadratic cost function
z(f ) =
1
2
|(QP+βR)f −Qp
reb
|
2
2
. (7.2)
Assuming that R is symmetric, the gradient of z(f ) equals
∇z(f ) = (QP+βR)
T
((QP+βR)f −Qp
reb
) (7.3)
= (P
T
Q
T
QP+βP
T
Q
T
R+βRQP+β
2
R
2
)f −(QP+βR)
T
Qp
reb
.
Therefore, any gradient based method for minimizing z(f ) would require calcu-
lation of matrix-vector products with P, P
T
, Q, and Q
T
in each iteration, i.e.,
standard tools from optimization theory for minimizing z(f ) in 7.2 are unnecce-
sarily inefficient and complicated to implement.
84 Relation between RIWFBP and least squares methods
In contrast to RIWFBP, most iterative methods are formulated as optimiza-
tion problems in which the error is measured in the projection domain rather than
the reconstruction domain, and does not involve an explicit reconstruction opera-
tor Q. If we could reformulate the RIWFBP method in similar terms, we would
be able to more easily compare it to competing methods, and to use tools from
optimization theory for better convergence analysis and improvements. Examples
of such improvements include automatic determination of step lengths for guaran-
teed convergence, and use of more advanced gradient based methods such as the
conjugate gradient method.
To find a quadratic cost function that is minimized by the RIWFBP method,
and measures the error in the projection domain, we have examined the possibility
to express the RIWFBP update step
f
k+1
= f
k
−αS(Q(Pf
k
−p
reb
) +βRf
k
), (7.4)
as a gradient descent step modified so that
f
k+1
= f
k
−αV∇z(f
k
) (7.5)
where V is symmetric and positive definite, and z(f ) is some quadratic function
to be minimized. Comparing (7.4) and (7.5), we see that the difference image in
RIWFBP must satisfy
V∇z(f ) = S(Q(Pf −p
reb
) +βRf ). (7.6)
Any gradient of a quadratic cost function is symmetric and non-negative. There-
fore, (7.6) is satisfied if and only if the matrix SQP can be written as VBP, where
BP is symmetric and non-negative. In such case, a factorization BP =
ˆ
P
T
W
ˆ
P
exist, where W is a symmetric non-negative weighting matrix, e.g. a rampfilter.
Then, the corresponding cost function would be
z(f ) =
1
2
|
ˆ
Pf −p
reb
|
2
W
+ Regularization term. (7.7)
However, the WFBP method employs a backprojection angle dependent nor-
malization (see Chapter 2), which makes it very difficult to express SQP as de-
scribed above. Therefore, it is highly unlikely that the RIWFBP method could be
expressed as in 7.5 for any quadratic cost function.
In the following sections, we present and evaluate the Weighted Least Squares
IFBP (WLS-IFBP) method, which is similar to the RIWFBP method but has an
update step that satisfies (7.5).
7.2 Weighted least squares IFBP (WLS-IFBP)
As shown in the previous section, it is difficult to analyze the RIWFBP method
in terms of a least squares problem. To obtain a method that is easier to analyze,
7.2 Weighted least squares IFBP (WLS-IFBP) 85
we consider the problem of finding a vector/reconstructed image f minimizing the
function
z(f ) =
1
2
|Pf −p
reb
|
2
W

N
¸
i=1
v
i
2
N
¸
j=1
d
ij
(f
i
−f
j
)
2
2
, (7.8)
which is a special case of the generally stated quadratic cost function proposed by
Delaney and Bresler [22].
In (7.8), | |
2
W
is defined as 'W, `, where ', ` denotes the Euclidean scalar
product. The matrix W equals W
Q
H, where H corresponds to a rampfilter,
and W
Q
corresponds to a downweighting of the most peripheral detector rows.
This is done to suppress artifacts associated with sharp transitions between voxels
receiving contributions and voxels not receiving contributions. The downweighting
function is the same as the one used for the WFBP method [100], and is given by
W
Q
(q) =





1, [q[ < Q
cos
2

π
2
|q|−Q
1−Q

, Q ≤ [q[ < 1
0, [q[ ≥ 1
, (7.9)
where q ∈ [−1, 1] is the detector row coordinate, normalized such that q = ±1
means the bottom and top boundaries of the detector respectively (see Fig. 2.10).
As will be demonstrated in the next section, a reasonable number for the parameter
Q is approximately 0.7. The coefficients v
i
are selected as the column sums of
W
Q
P in order to approximately preserve the local spatial resolution properties of
the RIWFBP method.
By differentiation of z(f ) and multiplication with the positive definite matrix
V = diag(1/

v
i
) S diag(1/

v
i
), (7.10)
we obtain the update step
f
k+1
= f
k
−α
k
V(
˜
Q(Pf
k
−p
reb
) +β
˜
Rf
k
). (7.11)
where
˜
Q = P
T
W, and the regularization operator is given by
˜
Rf =
N
¸
l=1

∂f
l


N
¸
i=1
v
i
2
N
¸
j=1
d
ij
(f
i
−f
j
)
2
2


e
l
=
N
¸
l=1


v
l
2
N
¸
j=1
d
lj
(f
l
−f
j
) −
N
¸
i=1
v
i
2
d
il
(f
i
−f
l
)


e
l
= ¦d
ij
= d
ji
, change of summation variables¦
=
N
¸
i=1
N
¸
j=1
d
ij
(v
i
+v
j
)
2
(f
i
−f
j
)e
i
. (7.12)
86 Relation between RIWFBP and least squares methods
Except for the outer parts of the field of view, variations between neighboring
values v
i
are small in most cases. Therefore, V ≈ S diag(1/v
i
) and V
˜
R is well
approximated by SR.
Comparing (7.11) and (7.4), we see that the WLS-IFBP method can be ob-
tained from the RIWFBP method by replacing SQ and SR with V
˜
Q and V
˜
R re-
spectively. Since the normalization in V
˜
Q allow certain angles to contribute more
than others (depending on the local amount of data redundancies), this operator
produces results corrupted by severe low-frequency artifacts when applied as a di-
rect reconstruction operator. Therefore, in order to reduce the number of required
iterations, the WLS-IFBP method is initialized with WFBP, i.e. f
0
= Qp
reb
.
Before continuing with experiments, we make one final remark regarding con-
vergence and choice of step lengths α
k
. The matrix
˜
R can explicitly be written as
˜
R =







¸
N
j=1
˜
d
1j


˜
d
11

˜
d
12

˜
d
1N

˜
d
21

¸
N
j=1
˜
d
2j


˜
d
22

˜
d
2N
.
.
.
.
.
.
.
.
.

˜
d
N1

˜
d
N2

¸
N
j=1
˜
d
Nj


˜
d
NN








(7.13)
where
˜
d
ij
= d
ij
(v
i
+ v
j
)/2. Since the symmetric matrix
˜
R is weakly diagonally
dominant, and thus non-negative, it can be factorized as
˜
R =
ˆ
R
T
ˆ
R. In the same
way, W can be factorized as W =
ˆ
W
T
ˆ
W. Thus, the cost function can also be
written as
z(f ) =
1
2
|
ˆ
W(Pf −p
reb
)|
2
2
+
β
2
|
ˆ
Rf |
2
2
=
1
2

ˆ
WP

β
ˆ
R

. .. .
=:A
f −

ˆ
Wp
reb
0

. .. .
=:b

2
2
. (7.14)
The step length α
k
for calculating f
k+1
can now be chosen as the global minimizer
of z(α) = z(f
k
+ αp
k
) where p
k
= −V∇z(f
k
). As shown in several textbooks,
e.g. Nocedal and Wright [71], this choice yields a globally convergent algorithm
for minimizing z(f ). In this case of a quadratic cost function, this step length can
be calculated analytically as
α
k
=
−p
T
k
∇z(f
k
)
p
T
k
A
T
Ap
k
. (7.15)
By reusing results, this automatic determination of α
k
can be made computa-
tionally inexpensive. However, it requires additional operations equivalent to one
iteration to be performed prior to starting the first iteration. To stay consistent
with the method and experiments presented in the previous chapter, we do not
use this automatic determination of α in the experiments that follow. Instead we
use a constant step length α
k
= α as determined in the previous chapter.
7.3 Experiments 87
a) WLS-IFBP, f
5
Backprojector: P
T
b) WLS-IFBP, f
5
Backprojector: B
S
Figure 7.1: WLS-IFBP reconstructions of the clock phantom using a) the Joseph
backprojector P
T
, and b) the “Standard” backprojector B
S
. Reconstruction pa-
rameters: Q = 0.7, β = 0.52, β
xy
= 1.0, β
z
= 1.5. Greyscale window ±20HU.
7.3 Experiments
7.3.1 Artifact reduction
We begin by comparing artifact suppressing abilities of the RIWFBP and the
WLS-IFBP using noise-free data. In any forward projection or backprojection
operation, the interpolation function must be adapted to the input sampling grid
in order to avoid DC-aliasing [12, 101]. The Joseph backprojector P
T
does not
satisfy this criterion, thus producing severe artifacts when used as backprojector
as shown in Fig. 7.1 a). To avoid these artifacts, we used the “standard back-
projector” B
S
, employing bilinear interpolation adapted to the detector sampling.
Fig. 7.1 b) shows the corresponding result obtained using B
S
. Note that results
obtained with the Joseph backprojector, or equivalently with symmetric forward
projector/backprojector pairs, can be improved for instance by modeling one de-
tector measurement as a mean value of several line integrals instead of just one.
The topic of improving the projection operator P is revisited in Chapter 9.
Fig. 7.2 b) shows a non-iterative WFBP reconstruction of the clock phantom.
As discussed in the previous chapter, two types of artifacts are present: (i) low-
frequency cone artifacts caused by the non-exactness of the reconstruction method,
and (ii) windmill artifacts showing up as alternating bright and dark regions near
sharp edges orthogonal to the z-direction of the reconstruction volume.
Fig. 7.2 c)-f) show results after one and two iterations with the RIWFBP and
WLS-IFBP methods. The regularization parameters previously determined to
give similar resolution properties for the RIWFBP and WFBP method were used
88 Relation between RIWFBP and least squares methods
(β = 0.52, β
xy
= 1.0, β
z
= 1.5). As expected, the RIWFBP method suppresses
cone artifacts slightly faster than the WLS-IFBP method. However, the difference
between the two methods is very small. After five iterations, when the two methods
have nearly converged, there are only small differences remaining in the close
vicinity of the outer solid spheres as shown in Fig. 7.3.
Fig. 7.4 shows WFBP, RIWFBP, and WLS-IFBP reconstructions for Q = 0.7
and Q = 1.0. For the iterative reconstructions, Q = 0.7 was used for initialization.
Obviously, using Q = 1.0 with non-iterative WFBP results in strong non-local
artifacts. Again, the RIWFBP and WLS-IFBP results look very similar. For both
methods, some new artifacts are introduced in the vicinity of the outer spheres as
the downweighting parameter Q is increased to 1.0 from 0.7. Also for Q = 1.0,
the artifacts are slightly more pronounced for the WLS-IFBP method than for the
RIWFBP method.
The errors σ
e1
(whole image) and σ
e2
(without edges) are shown in Fig. 7.5.
Only a very small reduction of the error over the whole image is observed. In this
particular experiment, the errors seem to get lower for Q = 1.0 than for Q = 0.3.
However, the relative difference is again very small, and error measurements of
head phantom reconstructions show the opposite relationship between the errors.
Over regions in the image, containing only low contrast structures (Ω
2
), an error
reduction from 5HU to 2HU is observed for all six reconstructions.
7.3.2 Spatial resolution and noise
Spatial resolution has been examined by measuring MTFs and SSPs with the
method described in Section 4.3, using edge phantoms of radius 20mm placed in
the isocenter, and at a distance of 150mm away from the isocenter.
MTF measurements for the RIWFBP and the WLS-IFBP method are shown
in Fig. 7.6. As demonstrated in the previous chapter, there is a large difference
in MTF depending on the distance to the isocenter. The MTFs for RIWFBP and
WLS-IFBP are nearly identical irrespective of distance to isocenter.
As shown in Fig. 7.7, also the SSPs for RIWFBP and WLS-IFBP are very
similar. In contrast to the xy-plane resolution, the distance to isocenter does not
seem to affect the SSPs much.
Fig. 7.8 shows noise measurements on clock phantom reconstructions for RI-
WFBP and WLS-IFBP at approximately equal spatial resolution. Again, there
are only small differences between the RIWFBP and WLS-IFBP reconstructions.
The choice of Q seem to be more important. In relation to WFBP, and depending
on the choice of Q, the iterative methods may decrease the noise level with 5%
(Q = 1.0), preserve the noise level (Q = 0.7), or increase the noise level with 5%
(Q = 0.3).
7.3 Experiments 89
a) Sampled phantom b) WFBP reconstruction
c) RIWFBP, f
1
d) WLS-IFBP, f
1
e) RIWFBP, f
2
f) WLS-IFBP, f
2
Figure 7.2: RIWFBP and WLS-IFBP reconstructions of the clock phantom.
Reconstruction parameters: Q = 0.7, β = 0.52, β
xy
= 1.0, β
z
= 1.5. Greyscale
window ±20HU.
90 Relation between RIWFBP and least squares methods
a) RIWFBP, f
5
b) WLS-IFBP, f
5
c) RIWFBP, f
5
d) WLS-IFBP, f
5
Figure 7.3: RIWFBP and WLS-IFBP reconstructions of the clock phantom after
five iterations. Reconstruction parameters: Q = 0.7, β = 0.52, β
xy
= 1.0, β
z
= 1.5.
Greyscale window ±20HU.
7.4 Discussions and conclusions
The RIWFBP method cannot easily be analyzed or understood with the tools from
optimization theory. However, as shown in Section 7.2, a small modification of the
method results in the WLS-IFBP method, which converges to the optimum of a
weighted least squares problem provided that reasonable step lengths α
k
are used.
The linear least squares formulation of the reconstruction problem makes it easy
to analytically determine the optimal choice of step length α, i.e. the step length
that minimizes the cost function along the search direction. Furthermore, the sim-
ilarity in formulation between WLS-IFBP and most other iterative reconstruction
methods potentially makes it easier to relate to and use ideas and techniques from
other methods.
From a practical point of view, in its current form, the WLS-IFBP does not
7.4 Discussions and conclusions 91
a) WFBP recon., Q = 0.7 b) WFBP recon., Q = 1.0
c) RIWFBP, f
5
, Q = 0.7 d) RIWFBP, f
5
, Q = 1.0
e) WLS-IFBP, f
5
, Q = 0.7 f) WLS-IFBP, f
5
, Q = 1.0
Figure 7.4: Comparison between WFBP, RIWFBP, and WLS-IFBP reconstruc-
tions for Q = 0.7 and Q = 1.0. All iterative reconstructions have been initial-
ized with a WFBP reconstruction using Q = 0.7. Regularization parameters:
β = 0.52, β
xy
= 1.0, β
z
= 1.5. Greyscale window ±20HU.
92 Relation between RIWFBP and least squares methods
0 5 10
49.8
49.9
50
50.1
50.2
50.3
Iteration number
R
M
S
E
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
RIWFBP, Q=0.3 RIWFBP, Q=0.7 RIWFBP, Q=1.0
0 5 10
0
1
2
3
4
5
Iteration number
R
M
S
E
,
w
i
t
h
o
u
t
e
d
g
e
s
(
H
U
)
WLS-IFBP,Q=0.3 WLS-IFBP,Q=0.7 WLS-IFBP,Q=1.0
Figure 7.5: σ
e
measurements on the clock phantom reconstructions in Fig. 7.2
and 7.3.
0 1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Frequency (lp/cm)
M
T
F
RIWFBP, f
5
, r = 0mm
WLS-IFBP, f
5
, r = 0mm
RIWFBP, f
5
, r = 150mm
WLS-IFBP, f
5
, r = 150mm
Figure 7.6: MTF measurements of the RIWFBP and the WLS-IFBP method in
the isocenter and 150mm off-center. The RIWFBP and WLS-IFBP curves overlap
both in the isocenter and off-center cases.
7.4 Discussions and conclusions 93
0 0.5 1 1.5 2 2.5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Distance (mm)
N
o
r
m
a
l
i
z
e
d
S
S
P
RIWFBP, f
5
, r = 0mm
WLS-IFBP, f
5
, r = 0mm
RIWFBP, f
5
, r = 150mm
WLS-IFBP, f
5
, r = 150mm
Figure 7.7: Normalized SSP measurements of the RIWFBP and the WLS-IFBP
method in the isocenter and 150mm off-center.
0 1 2 3 4 5 6 7 8 9 10
13.5
14
14.5
15
15.5
16
16.5
17
17.5
18
Iteration number
σ
n
,
(
H
U
)
RIWFBP, Q = 0.3
WLS-IFBP, Q = 0.3
RIWFBP, Q = 0.7
WLS-IFBP, Q = 0.7
RIWFBP, Q = 1.0
WLS-IFBP, Q = 1.0
Figure 7.8: σ
n
measurements made on clock phantom reconstructions.
94 Relation between RIWFBP and least squares methods
perform as good as the RIWFBP in terms of artifact reduction. However, consider-
ing the difference in how normalization is done in the two methods, this difference
in artifact reduction is smaller than expected.
For the projection operator used in this chapter, i.e. Joseph with only one
line integral per ray, using P
T
as a backprojector results in an increase of certain
artifacts. This is because the irradiation function (cf. Chapter 3) in this case
is one single sharp line per ray. In contrast, the “standard” backprojector B
S
,
modeling one ray with the triangle function implied from the linear interpolation,
has a regularizing effect on the result, thus avoiding increased artifact levels. As
shown in Chapter 9, the Joseph backprojector produces better results when more
line integrals are used for modeling one ray.
The choice of regularization parameters d
ij
and v
i
presented in this chapter
is ad hoc, and is primarily motivated by preservation of spatial resolution prop-
erties of RIWFBP and WFBP. Recent work by Shi and Fessler [91] in quadratic
regularization for statistically weighted reconstruction aim at, given a target PSF,
determine the optimal regularization parameters. Since this could provide means
of controlling the spatial resolution (e.g. to obtain uniform resolution), an inter-
esting future topic of research would be to investigate how such parameter choice
methods can be adapted to the methods presented here.
Chapter 8
Nonlinear regularization
8.1 Introduction
The methods studied in the previous sections are linear, i.e., any reconstructed
image f depends linearly on input data p. This property simplifies the analysis
and makes the results easy to predict: ignoring non-linearities in the acquisition,
the reconstruction of any object is then independent of local contrast level and
existence of other objects in the vicinity. Nevertheless, introduction of certain
types of nonlinearities in the reconstruction can improve image quality. One im-
portant example is statistical reconstruction, where the signal to noise ratio is
improved by taking the Poisson noise characteristics of input data into account
[21, 121, 126, 109, 98]. Many statistical reconstruction methods incorporate non-
linear regularization that reduces noise and spatial resolution for low contrast
while preserving high contrast structures. This is motivated by the fact that the
local signal to noise ratio is higher in high contrast structures than in low contrast
regions.
In this chapter, we examine the possibility to use nonlinear regularization in
RIWFBP and WLS-IFBP. The next section presents non-linearly regularized ver-
sions of the RIWFBP and WLS-IFBP methods, and so called potential functions,
employed to define the non-linearity of the regularization operator. This is followed
by a study of spatial resolution for different contrast levels. Having obtained an
understanding of the behavior of these methods for some parameter choices, we
study reconstructed images and noise measurements. We conclude by discussing
the results and presenting possible future research directions.
8.2 Description of the reconstruction method
To incorporate nonlinear regularization in the RIWFBP and WLS-IFBP methods,
we consider the following cost function
z
V
(f ) =
1
2
|Pf −p
reb
|
2
W

¸
N
i=1
v
i
2
¸
N
j=1
d
ij
V (f
i
−f
j
), (8.1)
96 Nonlinear regularization
which is a generalization of (7.8). The function V (f) is called potential function,
and describes how much different contrast levels should be penalized
1
. As in the
previous chapter, we differentiate z
V
(f ) and multiply with the positive definite
matrix V to obtain the update step
f
k+1
= f
k
−α
k
V(
˜
Q(Pf
k
−p
reb
) +β
˜
R
V
(f
k
)). (8.2)
Here, the regularization operator
˜
R
V
: R
N
→R
N
is given by
˜
R
V
(f ) =
N
¸
i=1
N
¸
j=1
d
ij
(v
i
+v
j
)
2
V

(f
i
−f
j
)e
i
. (8.3)
The derivative V

(f) is called influence function. Clearly, V

(f) and
˜
R
V
(f ) are
linear if and only if V (f) is a quadratic function.
Many different non-convex and convex potential functions have been suggested
for image reconstruction and restoration. Examples of non-convex potential func-
tions are the saturated quadratic function and the Geman and McClure potential
function [29]. Convex potential functions include the Huber penalty (see e.g. [21])
and the q-generalized Gaussian Markov Random field (q-GGMRF) penalty pro-
posed by Thibault et al. [109]. Since the sum of two convex functions is a convex
function, a convex potential function has the advantage that when it is used in
combination with a convex data term (such as a weighted square norm), the re-
sulting cost function also becomes convex whereby local minima become global
minima.
In the following, we will focus on the q-GGMRF function
V (f) =
[f[
p
1 +[f/c[
p−q
. (8.4)
The parameters p, q, and c are used to control the shape of the function, and
should satisfy 1 ≤ q ≤ p ≤ 2 and c > 0. In the left of Fig. 8.1, the q-GGMRF
functions are shown for different values of q. Obviously, for p = q = 2, this
function becomes a quadratic function. As the value of q is reduced from two
towards one, the potential function becomes increasingly v-shaped, and for q = 1,
an approximation of the Huber prior is obtained. As demonstrated in the right of
Fig. 8.1, the value of c determines where the intersection between the q-GGMRF
and the quadratic function should occur.
The q-GGMRF influence function is given by
V

(f) =
[f[
p−1
1 +[f/c[
p−q

p −

p −q
c
p−q

[f[
p−q
1 +[f/c[
p−q

sgn(f), (8.5)
and is illustrated in Fig. 8.2. As q decreases from two towards one, we observe that
the slope of the influence function becomes steeper for low contrasts and flatter
for high contrasts.
1
The details of the regularization term are related to the statistical description of an image
as a stochastic vector belonging to a Gibbs distribution. A comprehensive text on this topic is
found in [63].
8.2 Description of the reconstruction method 97
−40 −20 0 20 40
0
200
400
600
800
Contrast (HU)
P
e
n
a
l
t
y
Quadratic (q = p = 2)
p = 2.0, q = 1.7, c = 20
p = 2.0, q = 1.3, c = 20
p = 2.0, q = 1.0, c = 20
−40 −20 0 20 40
0
200
400
600
800
Contrast (HU)
P
e
n
a
l
t
y
Quadratic (q = p = 2)
p = 2.0, q = 1.3, c = 10
p = 2.0, q = 1.3, c = 20
p = 2.0, q = 1.3, c = 30
Figure 8.1: q-GGMRF potential functions for different values of q and c.
−40 −30 −20 −10 0 10 20 30 40
−40
−30
−20
−10
0
10
20
30
40
Contrast (HU)
I
n

u
e
n
c
e
(
H
U
)
Quadratic (q = p = 2)
p = 2.0, q = 1.7, c = 20
p = 2.0, q = 1.3, c = 20
p = 2.0, q = 1.0, c = 20
Figure 8.2: q-GGMRF influence functions for different values of q.
98 Nonlinear regularization
For the RIWFBP method, the new update formula becomes
f
k+1
= f
k
−αS(Q(Pf
k
−p
reb
) +βR
V
f
k
), (8.6)
where
R
V
(f ) =
N
¸
i=1
N
¸
j=1
d
ij
V

(f
i
−f
j
)e
i
. (8.7)
8.3 Experiments
8.3.1 Spatial resolution versus local contrast
In the previous chapter, for linear regularization, the results of WLS-IFBP and
RIWFBP were found to be very similar. Assuming this similarity to be valid
also for non-linear regularization we limit the experiments in this chapter to the
WLS-IFBP method.
To assess the spatial resolution for different contrast levels, we have used the
solid sphere phantom from previous chapters, but with different linear attenuation
factors. For practical reasons, and also to prevent the influence of nonlinear partial
volume effects, different linear attenuation factors were obtained by linearly scaling
the projection data. The different contrast levels studied were 10HU, 20HU, 40HU,
60HU, 80HU, 100HU, 150HU, 200HU, 300HU, 400HU, and 500HU.
In all experiments presented throughout this chapter, projection data have
been pre-filtered as described in Chapter 5. Furthermore, the coefficients d
ij
have
been selected so that the nonlinear regularization described here perfectly coincides
with the linear regularization from previous chapters when p = q = 2 (β
xy
= 1 and
β
z
= 1.5). One could argue that input data should be pre-processed as little as
possible, and that d
ij
should be selected as inverse distances to the center voxel,
which is the traditional choice. However, our choice simplifies comparison with
results from previous chapters.
With nonlinear regularization, resolution measurements become structure de-
pendent. For instance, an edge based method yields a different result than a line
based method. Therefore, in the following one should keep in mind that the pre-
sented MTFs and SSPs are to be interpreted as characterizations of edge responses.
The fact that MTF and SSP measurements are phantom dependent for nonlinear
reconstruction methods is further discussed in the next section.
As the regularization parameter q is decreased from two towards one, more
iterations are required for convergence. This is demonstrated in Fig. 8.3, show-
ing difference norms log
10
(|f
k+1
− f
k
|/C) and SSP full width at half maximum
(FWHM) as a function of iteration number. For low values of q, the difference
norm is still decreasing after 50 iterations. Some cross-plane resolution contribu-
tions seem to remain.
8.3 Experiments 99
0 10 20 30 40
−5
−4
−3
−2
−1
0
Iteration number
l
o
g
1
0
(
|
f
k
+
1

f
k
|
)
Quadratic
p = 2.0, q = 1.7, c = 20
p = 2.0, q = 1.3, c = 20
p = 2.0, q = 1.0, c = 20
0 10 20 30 40 50
1
1.2
1.4
1.6
Iteration number
F
W
H
M
(
m
m
)
Quadratic
p = 2.0, q = 1.7, c = 20
p = 2.0, q = 1.3, c = 20
p = 2.0, q = 1.0, c = 20
Figure 8.3: Left: the difference norm log
10
(|f
k+1
− f
k
|/C) plotted against the
iteration number for central images (z = 0mm) of resolution phantom reconstruc-
tions with different choices of regularization parameters q. Right: SSP Full Width
at Half Maximum (FWHM) as a function of the iteration number.
Fig. 8.4 shows ρ
50%
values
2
as a function of contrast level after 50 iterations.
Since a quadratic potential function corresponds to a linear reconstruction method,
the ρ
50%
for the quadratic potential function is constant with respect to the con-
trast level. To the left, the value of the regularization parameter q is varied. As q
decreases, a slight reduction of ρ
50%
for low contrasts and a slight increase of ρ
50%
for high contrasts is observed. In this particular case, where c = 20, the intersec-
tion of the ρ
50%
curves is located at a contrast level of approximately 30HU. To
the right in Fig. 8.4, ρ
50%
curves are shown for different values of c. By increasing
c, the intersection of the ρ
50%
curves moves towards higher contrast levels.
Fig. 8.5 shows the same plots as Fig. 8.4 but after ten iterations. Hardly
any differences can be observed, i.e., in terms of in-plane resolution, convergence
seems to be obtained in approximately ten iterations.
Fig. 8.6 shows Full Width at Half Maximum (FWHM) values of the SSP as a
function of contrast level after 50 iterations. Here, reducing q has a greater impact
than it has on in-plane resolution (MTF ρ
50%
). While, for q = 1.3 and c = 20,
the MTF ρ
50%
increases with approximately 10%, a decrease of more than 30% is
observed for the SSP FWHM at contrast levels above 300HU.
Another observation in Fig. 8.6 is the unexpected behavior of the SSP FWHM
curve for q = 1.0. Initially, the curve drops rapidly from 1.50mm at 10HU to
2
The ρ
50%
value of an MTF is the frequency, at which the amplitude has dropped to 50%.
100 Nonlinear regularization
0 100 200 300
3.4
3.6
3.8
4
Contrast (HU)
ρ
5
0
%
(
l
p
/
c
m
)
Quadratic
q = 1.7, c = 20
q = 1.3, c = 20
q = 1.0, c = 20
0 100 200 300
3.4
3.6
3.8
4
Contrast (HU)
ρ
5
0
%
(
l
p
/
c
m
)
Quadratic
q = 1.3, c = 10
q = 1.3, c = 20
q = 1.3, c = 30
Figure 8.4: ρ
50%
of the MTF for different contrast levels and regularization
parameters q and c. The number of iterations equals 50.
0 100 200 300
3.4
3.6
3.8
4
Contrast (HU)
ρ
5
0
%
(
l
p
/
c
m
)
Quadratic
q = 1.7, c = 20
q = 1.3, c = 20
q = 1.0, c = 20
0 100 200 300
3.4
3.6
3.8
4
Contrast (HU)
ρ
5
0
%
(
l
p
/
c
m
)
Quadratic
q = 1.3, c = 10
q = 1.3, c = 20
q = 1.3, c = 30
Figure 8.5: ρ
50%
of the MTF for different contrast levels and regularization
parameters q and c. The number of iterations equals 10.
8.3 Experiments 101
0 100 200 300
0.8
0.9
1
1.1
1.2
1.3
1.4
Contrast (HU)
F
W
H
M
(
m
m
)
Quadratic
q = 1.7, c = 20
q = 1.3, c = 20
q = 1.0, c = 20
0 100 200 300
0.8
0.9
1
1.1
1.2
1.3
1.4
Contrast (HU)
F
W
H
M
(
m
m
)
Quadratic
q = 1.3, c = 10
q = 1.3, c = 20
q = 1.3, c = 30
Figure 8.6: SSP FWHM for different contrast levels and regularization parame-
ters q and c. The number of iterations equals 50.
around 0.8mm at 200HU. As the contrast further increases, the FWHM starts
to increase, and levels off at approximately 1.0 mm (not visible in the figure).
SSPs for this case are shown in Fig. 8.7. For a contrast level of 200HU and
q = 1.0, it seems like the combination of much smoothing for low contrasts and
very little smoothing at high contrasts helps to sharpen the SSP more than in
the non-regularized case. As the contrast level increases to 500HU, the relative
amount of low contrast smoothing becomes smaller, and the SSP approaches that
of the non-regularized method.
As shown already in Fig. 8.3, the number of iterations needed for convergence
of cross-plane resolution is high for values of q close to one. This is verified in Fig.
8.8, showing the same plots as Fig. 8.6 but for 50 iterations. Although the plots
are similar in both cases, the improvements are less pronounced after 10 iterations
than after 50. For q = 1.3, c = 20, the previously observed 30% decrease of FWHM
after 50 iterations is approximately 20% after 10 iterations.
8.3.2 Edge versus surface based resolution measurements
Most expressions used for characterizing spatial resolution in an imaging system
are based on the assumption that the system is approximately linear. Under
this assumption, the spatial resolution properties are fully characterized by the
spatially variant or invariant point spread function (PSF). Derivations of edge
based measurement methods rely on the linearity to show how the PSF can be
calculated from an edge response.
102 Nonlinear regularization
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Distance from slice center (mm)
S
S
P
Quadratic
No regularization
q = 1.0, c = 20, 200HU
q = 1.0, c = 20, 500HU
Figure 8.7: Comparison of SSPs for quadratic regularization, no regularization,
and strongly nonlinear regularization (q = 1.0, c = 20) for the two contrast levels
200HU and 500HU. In all cases, the number of iterations equals fifty.
0 100 200 300
1
1.1
1.2
1.3
1.4
1.5
Contrast (HU)
F
W
H
M
(
m
m
)
Quadratic
q = 1.7, c = 20
q = 1.3, c = 20
q = 1.0, c = 20
0 100 200 300
1
1.1
1.2
1.3
1.4
1.5
Contrast (HU)
F
W
H
M
(
m
m
)
Quadratic
q = 1.3, c = 10
q = 1.3, c = 20
q = 1.3, c = 30
Figure 8.8: SSP FWHM for different contrast levels and regularization parame-
ters q and c. The number of iterations equals 10.
8.3 Experiments 103
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
Distance from slice(mm)
S
S
P
WFBP, Edge
WFBP, Surface
f
10
, q = 1.3, Edge
f
10
, q = 1.3, Surface
Figure 8.9: Comparison between edge based and surface based SSP measure-
ments in presence of nonlinear regularization. The number of iterations equal
ten.
An imaging system using nonlinear regularization as presented in this chapter
cannot be assumed to be “approximately linear”. Therefore, the measured spatial
resolution for such a system depends on the type of structure studied, and conse-
quently, in the previous section it was pointed out that the presented MTFs and
SSPs should be interpreted as characterizations of edge responses.
We illustrate the dependence of structure with an example. Fig. 8.9 shows
SSPs for WFBP and WLS-IFBP(f
10
, p = 2.0, q = 1.3, c = 20) reconstructions of
the two phantoms described in Section 4.3. One of these is a solid sphere referred
to as the edge phantom, and the other one is a 0.1mm thick shell referred to as
the surface phantom. Both phantoms have a density corresponding to 1000HU
above air. For the WFBP reconstructions, the SSP measurements differ very
little from each other. However, for the nonlinear reconstructions, the surface
SSP is clearly wider than the edge SSP. This difference is explained by (i) thick-
ness dependent dampening of the reconstructed profile of the thin shell phantom
(0.1mm⇒ 1000HU→ 80HU), and (ii) the difference in shape between the profiles
of the shell and edge phantoms. The non-normalized peak of the line SSP is 80HU.
For this contrast level, the edge FWHM is reduced by 17% and the line FWHM
is reduced by 12%.
104 Nonlinear regularization
8.3.3 Artifacts and noise
Fig. 8.10 b) shows an axial image of a WFBP head phantom reconstruction from
noise-free data. The image is corrupted by streak artifacts plus a combination of
cone-artifacts, wind-mill artifacts, and partial volume artifacts originating from
the high contrast structures in the inner ear [129]. As shown in Fig. 8.10 c),
after ten iterations with WLS-IFBP using quadratic regularization, the strong
artifacts originating from the inner ear have been alleviated, but are still visible
in this particular greyscale window. In Fig. 8.10 d), attained with nonlinear
regularization (p = 2.0, q = 1.3, c = 20), the artifacts are even more alleviated. In
addition, the nonlinear regularization also improve the image in terms of spatial
resolution.
These observations are confirmed by the σ
e
measurements plotted in Fig. 8.11.
To the left, the measurements stem from the whole image. Depending on the
choice of regularization parameter c, this σ
e
value is reduced from 102HU for the
quadratic (linear) case to somewhere between 95HU and 96HU. The right part of
this figure shows σ
e
calculated on low contrast regions. For all iterative methods,
there is a reduction of this value from 10.4HU to just below 9HU. The reduction
is slightly higher for the nonlinearly regularized reconstructions, an effect which
seems rather independent of the parameter c.
For any linear reconstruction method, the pixel noise is inversely proportional
to the square root of the number of photons, i.e., increasing the number of photons
with a factor of four results in a reduction in pixel noise by a factor of two [46].
However, for nonlinear methods, this does no longer hold true. To examine the
noise properties for the nonlinearly regularized WLS-IFBP, we have studied head
phantom reconstructions for different numbers of non-attenuated photons, and
for different values of the regularization parameter c. To measure the maximum
noise levels, measurements have been restricted to the center of the reconstructed
objects.
Fig. 8.12 shows σ
n

nq
for noise levels corresponding to ¦4
k
10
6
¦
2
k=−2
non-
attenuated photons. Here, σ
nq
are noise measurements for the quadratically regu-
larized WLS-IFBP method, for which the above presented proportionality holds.
In relation to the linearly regularized WLS-IFBP method, the nonlinearly regu-
larized WLS-IFBP suppresses noise at low noise levels, and amplifies noise at high
noise levels. The transition point, i.e., where the nonlinear and linear methods
give similar results, can be controlled by changing the regularization parameter c.
Fig. 8.13 shows three axial images of the physical head phantom reconstructed
with WFBP, linear WLS-IFBP, and non-linear WLS-IFBP. The WFBP recon-
struction is corrupted by windmill artifacts. As shown in Fig. 8.13 b), the linear
WLS-IFBP method gives a better result with less pronounced wind-mill artifacts.
The nonlinear WLS-IFBP in Fig. 8.13 c) further reduces the amplitude of these
artifacts, as well as improves the cross-plane resolution as can be readily seen in
the encircled areas.
8.3 Experiments 105
a) Sampled phantom b) WFBP reconstruction
c) WLS-IFBP, f
10
, p = q = 2 d) WLS-IFBP, f
10
,
p = 2.0, q = 1.3, c = 20
Figure 8.10: Axial images of noise-free head phantom reconstructions. Greyscale
window:−25HU,25HU.
106 Nonlinear regularization
0 5 10
95
96
97
98
99
100
101
102
Iteration number
R
M
S
E
,
w
h
o
l
e
i
m
a
g
e
(
H
U
)
p = 2.0, q = 2.0 q = 1.3, c = 10 q = 1.3, c = 20
0 5 10
8
8.5
9
9.5
10
10.5
Iteration number
R
M
S
E
,
l
o
w
c
o
n
t
r
a
s
t
r
e
g
i
o
n
s
(
H
U
)
q = 1.3, c = 30 q = 1.3, c = 40
Figure 8.11: WLS-IFBP σ
e
values measured on head phantom reconstructions.
10
5
10
6
10
7
0.9
0.95
1
1.05
1.1
1.15
1.2
Number of unattenuated photons
σ
n
/
σ
n
q
σ
nq
= 22.6HU
11.1HU 5.6HU 2.8HU 1.4HU
q = 1.3, c = 10
q = 1.3, c = 20
q = 1.3, c = 30
q = 1.3, c = 40
Figure 8.12: σ
n

nq
against the number of unattenuated photons. σ
nq
are noise
measurements for WLS-IFBP with quadratic regularization. The noise measure-
ments were made on central parts of the phantom. In all cases, the number of
iterations are 10.
8.3 Experiments 107
a) WFBP reconstruction b) WLS-IFBP, f
10
, p = q = 2.0
c) WLS-IFBP, f
10
, q = 1.3, c = 20
Figure 8.13: Axial images of head phantom reconstructions. Greyscale window
±75HU.
108 Nonlinear regularization
Fig. 8.14 shows axial images of physical thorax phantom reconstructions.
Again, we observe the presence of wind-mill artifacts in the WFBP reconstruc-
tion, and how the amplitude of these to a certain extent can be reduced with the
linear WLS-IFBP method. For the non-linear WLS-IFBP method, two values of β
were used. As shown in Fig. 8.14 c), the nonlinear WLS-IFBP with β = 0.6 does
not improve the result much over the linear WLS-IFBP. However, by increasing
β to 0.9, further artifact reduction can be obtained. Although the high contrast
spatial resolution is worse in Fig. 8.14 d) than Fig. 8.14 c), it is still better than
for the linear methods.
8.4 Discussions and conclusions
The results in this chapter indicate that nonlinear regularization can be used for
improving results of WLS-IFBP reconstruction with respect to wind-mill artifact
reduction and high contrast spatial resolution, without increasing noise in low
contrast regions. This is an effect of using relatively low penalties for high contrast
structures, and relatively high penalties for low contrast structures, resulting in
sharpening of high contrast structures and softening of low contrast structures.
Another way of describing this is to view the nonlinearly regularized reconstruction
as using two reconstruction kernels
3
: one sharp kernel for high contrast structures,
and one soft kernel for low contrast structures.
In terms of artifact and spatial resolution, most improvement take place within
the first ten iterations. However, for potential functions that deviate much from
a quadratic function, e.g. for q close to one, more iterations are needed until no
further improvement can be observed. This is partially explained by the constant
step length α = 0.7 and the low-pass filtering S introduced in Section 5.1. As
illustrated by the simple one-dimensional example in Fig. 8.15, the optimal step
length α depends on the shape of the cost function. Fig. 8.15, as well as prelimi-
nary experiments, further suggests that the convergence rate could be improved by
using larger step lengths for high contrast structures than in low contrast regions.
Another possibility for improving the convergence rate is to use a more advanced
optimization method such as the preconditioned conjugate gradient method or
ordered subsets acceleration.
Fig. 8.15 illustrates a problem associated with step lengths that are constant
with respect to iteration number. While this works fine for a quadratic potential
function (see Section 5.1), an infinite cycle of sub-optimums might occur for a non-
quadratic potential function. A convergent method is obtained if the step lengths
α
k
satisfy the so called Wolfe conditions [71]. Such step lengths can be found by
performing a so called line search in each iteration. A practical line search method
3
A reconstruction kernel is a rampfilter convolved with a resolution modifying filter as de-
scribed in Chapter 5.
8.4 Discussions and conclusions 109
a) WFBP reconstruction
b) WLS-IFBP, f
10
, p = q = 2.0
c) WLS-IFBP, f
10
, q = 1.3, c = 50, β = 0.6
d) WLS-IFBP, f
10
, q = 1.3, c = 50, β = 1.0
Figure 8.14: Axial images of thorax phantom reconstructions. Greyscale win-
dow: ±150HU.
110 Nonlinear regularization
−50 0 50
0
500
1000
1500
f
1
(x) = x
2
/2, α = 1.9
x
f
1
(
x
)
−50 0 50
0
500
1000
1500
f
2
(x) =
|x|
p
1+|x/c|
p−q
, α = 1.9
x
f
2
(
x
)
−50 0 50
0
500
1000
1500
f
1
(x) = x
2
/2, α = 0.8
x
f
1
(
x
)
−50 0 50
0
500
1000
1500
f
2
(x) =
|x|
p
1+|x/c|
p−q
, α = 0.8
x
f
2
(
x
)
Figure 8.15: One-dimensional example of gradient descent optimization of two
functions f
1
(x) and f
2
(x) using two different constant step lengths α. The param-
eters for f
2
are p = 2, q = 1, and c = 20. In all cases, the iterative method has
been initialized with x
0
= −50.
requiring no forward projections or backprojections has been proposed by Fessler
and Booth [25].
The q-GGMRF potential function adds a number of regularization parameters
to the cost function. These parameters must be carefully adjusted for every type
of examination that the CT system is intended for. In order not to enhance noise,
the maximum noise level must be estimated and used for properly choosing the
regularization parameter c. This is an important difference between the methods
studied here and statistical methods. In statistical reconstruction methods, the
noise estimation problem is avoided by downweighting noisy data contributions
relative to the regularization term, leading to blurring along directions orthogonal
to the ray projection.
From a clinical point of view, artifact suppression and improved spatial resolu-
8.4 Discussions and conclusions 111
tion of the nonlinearly regularized WLS-IFBP method will highlight low contrast
structures otherwise distorted by artifacts or located close to bone. However, it
remains unclear whether the observed noise reduction in low contrast regions im-
proves detection of low contrast objects. A human observer study is needed to
draw any further conclusions in this matter, (see e.g. Borsdorf et al. [4]).
112 Nonlinear regularization
Chapter 9
Data acquisition modeling
9.1 Irradiation functions
As discussed in Chapter 3, iterative reconstruction offers the possibility to ac-
curately model the data acquisition process in the forward projection P. Still,
throughout previous chapters we have employed the single ray Joseph method
1
,
which is a rather poor model of the physical projection event.
We recall from Section 3.3.1 that a measurement p
i
acquired from a recon-
struction result f can be expressed as
p
i
=

R
3
w
i
(r)f
c
(r)dr, (9.1)
where f
c
: R
3
→ R is the continuous function obtained by interpolation with the
basis function b : R
3
→ R
2
. The functions w
i
: R
3
→ R are so called irradiation
functions that define how f
c
contributes to each of the measurements p
i
respec-
tively, and can be used to incorporate accurate models of the data acquisition in
the forward projection operator P.
For the single ray Joseph method, the irradiation function w
i
for a certain
measurement p
i
is a single Dirac line intersecting the source and detector of this
specific measurement. This is equivalent to assuming that the source and detector
are point shaped. The fact that the gantry is rotating during acquisition is also
ignored.
In the following section, we investigate the effects of improving the projection
operator P by using irradiation functions consisting of multiple Dirac lines, sim-
1
We use the words “single ray” to emphasize that the Joseph method appearing in previous
chapters calculates one single line integral per measurement.
2
To make (9.1) applicable to Joseph forward projection, we must allow the basis functions
to be view-dependent as illustrated in Fig. 3.9. This should not have dramatic effects on the
reconstruction results, since the Joseph method generates projection data that are very similar to
the method by K¨ohler and Turbell [51], that employs a view-independent tri-linear basis function
[110].
114 Data acquisition modeling
ulating the spatial extent of the source and detector, as well as gantry rotation.
This is followed by a presentation of a projection method based on the Joseph
method, but calculating strip integrals instead of line integrals.
9.2 Multiple ray forward projection
9.2.1 An improved acquisition model
To get most out of any modeling attempt, the input data should be preprocessed
as little as possible, i.e., rebinning should not be a preprocessing step but rather be
put inside the iterative loop, or entirely avoided. However, previous experiments in
[102] have shown that putting the rebinning inside the iterative loop leads to strong
overshoots, presumably caused by complete suppression of certain high frequencies
in the rebinning. By formulating a rampfilter weighted least squares problem in
the cone-beam data domain, it could be possible to avoid rebinning entirely, but
this has not been investigated so far. Instead, we accept the low-pass filtering
caused by rebinning, and focus on improving the reconstruction by modeling the
low-pass filtering in the acquisition process.
We start with a description of the improved cone-beam geometry model, and
then explain how this model can be applied to the rebinned geometry. We assume
that the source (the focal spot) is shaped like a rectangle sticking to the cone-
shaped anode as shown in Fig. 9.1 and 9.2. The center r
SC
of the rectangular
source, and the unit vectors ˆr
SW
and ˆr
SL
pointing along the main directions of
the rectangle, are given by
r
SC
(α) =

R
S
cos α, R
S
sin α,
αP

T
,
ˆr
SW
(α) =

−sin α, cos α, 0

T
, and
ˆr
SL
(α) =

−cos αsin η, −sin αsin η, cos η

T
. (9.2)
The so called anode angle η is the angle between ˆr
SL
and the z-axis. Any point
on the rectangle can now be expressed as a convex combination of the four vectors
defining the corners of the rectangle: r
SC
±(w
S
/2)ˆr
SW
, and r
SC
±(l
S
/2)ˆr
SL
.
We use a rectangular model also for the detector elements. When projected
onto the xy-plane, the detector elements face the source as shown in Fig. 9.1.
Furthermore, the detector elements are aligned to the z-axis as shown in Fig. 9.2.
In the same way as for the source, the detector elements are described by
r
DC
(α, β, q) = r
SC
(α) +R
SD

−cos(α +β), −sin(α +β),
qh
2R
S

T
ˆr
DW
(α, β) =

sin(α +β), −cos(α +β), 0

T
ˆr
DH
=

0, 0, 1

T
. (9.3)
Let d
β
and d
q
be factors describing the relation between the detector sampling
distances and the actual width and height of the detector elements. The four
9.2 Multiple ray forward projection 115

SC

DC
v
x

SD
o
ß
´
SW
´
DW
X-ray source Detector element
R
SD
R
S
l
S
]cos¦n-n)]
w
S
R
SD
a
ß
A
ß
Figure 9.1: Illustration of the shape of the source and detector, and involved
unit vectors in the xy-plane. Note that the sizes of the source and detectors are
exaggerated.

SC

DC
z
x

SL

DH
X-ray source Detector element

SD

R
SD
a
q

q
h
2 R
S
Figure 9.2: Illustration of the shape of the source and detector, and involved
unit vectors in the xz-plane. Note that the sizes of the source and detectors are
exaggerated.
116 Data acquisition modeling
0 1 2
l 0
1
2
3
4

SW

SL
m0 1 2
n0
1
2
3

DW

DH
Source, N
SW
÷3, N
SL
÷5
w
S
l
S
w
D
h
D
Detector, N
DW
÷3, N
DH
÷4
Figure 9.3: Arrangement of points for a multiple ray forward projection model.
vectors pointing out the corners of a detector element are then given by r
DC
±
(w
D
/2)ˆr
DW
and r
DC
±(h
D
/2)ˆr
DH
, where the width w
D
and height h
D
are given
by
w
D
= R
SD
d
β
Δ
β
, and h
D
=
R
SD
d
q
Δ
q
h
2R
S
. (9.4)
Having defined the rectangles for the source and detector, the next step is
to divide these rectangles into multiple points, and to calculate measurements as
mean values of line integrals between all source points and all detector points. Let
N
SW
and N
SL
be the number of source points along the ˆr
SW
and ˆr
SL
respectively.
The points used for simulating the spatial extent of the source are then given by
r
S
[i, k, l] = r
SC

i
)+
+
w
S
N
SW


N
SW
−1
2
+k

ˆr
SW
+
l
S
N
SL


N
SL
−1
2
+l

ˆr
SL
,
k = 0, ..., N
SW
− 1, l = 0, ..., N
SL
− 1. (9.5)
Similarly, the points used for simulating the spatial extent of the detector are given
by
r
D
[i, m, n] = r
DC

i
, β
i
, q
i
)+
+
w
D
N
DW


N
DW
−1
2
+m

ˆr
DW
+
h
D
N
DH


N
DH
−1
2
+n

ˆr
DH
,
m = 0, ..., N
DW
− 1, n = 0, ..., N
DH
− 1. (9.6)
The arrangements of the points on the source and detector elements are illustrated
in Fig. 9.3.
9.2 Multiple ray forward projection 117
With Γ
iklmn
denoting the line intersecting the points r
S
[i, k, l] and r
D
[i, m, n],
the ith measurement can be expressed as
p
i
=
1
N
SW
N
SL
N
DW
N
DH
N
SW
−1
¸
k=0
N
SL
−1
¸
l=0
N
DW
−1
¸
m=0
N
DH
−1
¸
n=0

Γ
iklmn
f
c
(r)dl. (9.7)
Using the same technique as for the source and detector modeling, the acqui-
sition model can be further improved by taking the gantry rotation into account.
We let N
GR
denote the number of sub-angles used to simulate this effect.
When applying this model in the rebinned geometry, we need to make sure that
the unit vectors ˆr
SW
, ˆr
SL
, ˆr
DW
, and ˆr
DH
are computed correctly. This is done
by calculating (α,β) from (θ,p) by using the rebinning equations β = sin
−1
(p/R
S
)
and α = θ −β, and applying (9.2) and (9.3).
9.2.2 Experiments
We have studied spatial resolution for different choices of the forward projection
parameters N
SL
, N
SW
, N
DH
, N
DW
, and N
GR
. In all experiments, input data
have been generated as described in Chapter 4, using the scanning parameters in
Table 4.1. Two resolution phantoms were used: one located in the iso-center and
one located 150mm off-center for measuring off-center radial resolution, as well as
azimuthal resolution.
For more accurate modeling of any of the previously presented entities (e.g.
length of the source), its corresponding N-value (e.g. N
SL
) should be set to a value
larger than one. The higher the N-value is, the more accurate the model becomes.
Unfortunately, since the total cost of a multiplication with P is directly propor-
tional to the product N
SL
N
SW
N
DH
N
DW
N
GR
, high N-values quickly
leads to prohibitively high computational costs. In the experiments, the N-values
have been empirically selected such that any further increase in N-value yields an
improvement that is very small in relation to the already obtained improvement.
For example, N
DH
= 3 reduces the SSP FWHM from 1.53mm to 1.29mm. If the
value of N
DH
is increased to four, the SSP FWHM remains unchanged at 1.29mm.
In contrast to previous chapters, we present results obtained not only with
the “standard backprojector”
3
B
S
, but also with the transpose P
T
of the forward
projector as suggested by the derivation of the gradient descent method in Section
7.2. As illustrated in Fig. 9.4, due to the mismatched size of the interpolation
function, DC-aliasing artifacts may occur when using P
T
corresponding to the
Joseph method with only one ray per measurement.
Interestingly, DC-aliasing artifacts during backprojection can be reduced by
increasing the number of rays per measurement using the techniques described in
the previous section. This is illustrated in Fig. 9.5. The WFBP reconstruction in
3
The “standard backprojector” B
S
refers to a backprojector employing linear/bilinear inter-
polation adapted to the detector sampling grid. We use the word “standard” because this is the
backprojector most frequently found in the literature.
118 Data acquisition modeling

t
t

=0
a) Standard backproiector

. Interpolation:

=0
b) Joseph backproiector

. Interpolation:

=
2

3
2

2

Figure 9.4: Illustration of DC-aliasing during backprojection. Since the interpo-
lation function in b) is not adapted to the projection data sampling distance, false
high frequencies occur in the backprojected images.
Fig. 9.5 a) is completely free from artifacts originating from interpolations in the
backprojection. In Fig. 9.5 b) we see that WLS-IFBP with B
S
as backprojector
does add some high frequency artifacts. However, the amplitude of these artifacts
is low, and can be further reduced by increasing the sampling density of the re-
construction volume [102]. When B
S
is replaced with P
T
, severe artifacts occur
as shown in Fig. 9.5 c). These DC-aliasing artifacts occur mainly because the in-
terpolation function in the z-direction is too small: the detector sampling distance
is hΔ
q
/2 = 1.2mm, and the size of the interpolation function is determined only
with respect to the much smaller voxel volume sampling distance Δ
z
= 0.5mm.
As the number of Dirac lines is increased from one to three in the detector height
direction, the size of the interpolation function becomes more adapted to the de-
tector sampling distance, and the amplitude of the DC-aliasing artifacts is reduced
as shown in Fig. 9.5 d). Fig. 9.5 e) and f) show that this type of artifacts can be
further reduced by adding more rays corresponding to the detector width and the
spatial extent of the source. In the following experiments, modeling of the detec-
tor element height with N
DH
= 3 will always be used for suppressing DC-aliasing
artifacts.
We begin with studying cross-plane resolution in the iso-center, and how it is
affected by modeling of the source length N
SL
and detector element height N
DH
in the forward projection model. It is reasonable to assume that the other param-
eters N
SW
, N
DW
, and N
GR
have negligible impact on the cross-plane resolution.
9.2 Multiple ray forward projection 119
a) WFBP reconstruction b) WLS-IFBP, f
10
, B
S
N
SL
= 1, N
SW
= 1
N
DH
= 1, N
DW
= 1
c) WLS-IFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 1, N
DW
= 1
d) WLS-IFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 3, N
DW
= 1
e) WLS-IFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 3, N
DW
= 2
f) WLS-IFBP, f
10
, P
T
N
SL
= 3, N
SW
= 2
N
DH
= 3, N
DW
= 2
Figure 9.5: Axial images of MTF phantom reconstructions for different back-
projectors and number of rays per measurement. Greyscale window (-1025HU,
975HU). In all cases, the N
GR
equals one.
120 Data acquisition modeling
0 5 10 15
1.25
1.3
1.35
1.4
1.45
1.5
1.55
1.6
Iteration number
S
S
P
F
W
H
M
(
m
m
)
B
S
, N
SL
= 1
B
S
, N
SL
= 3
P
T
, N
SL
= 1
P
T
, N
SL
= 3
Figure 9.6: SSP FWHM curves as a function of iteration number for WLS-IFBP.
Both B
S
and P
T
were used as backprojectors. In all plots, the detector height is
taken into account with N
DH
= 3. Except for N
SL
, other multiple ray parameters
equal one.
Therefore, these parameters have all been set to one. As seen in Fig. 9.6, the
SSP FWHM stops improving after approximately 10 iterations. At this point, the
FWHM has decreased from 1.53mm to around 1.43mm with B
S
, and 1.29mm with
P
T
. Only small differences are observed when N
SL
is increased from one to three.
Unexpectedly, in the case of the P
T
backprojector, increasing N
SL
from one to
three leads to a slightly increased FWHM.
Fig. 9.7 shows the SSPs from Fig. 9.6 after ten iterations. In addition to the
previous observations, we see that using B
S
instead of P
T
, as well as increasing
N
SL
from one to three lead to an amplification of the overshoots.
In-plane resolution can be divided into radial resolution in the iso-center, and
radial and azimuthal (angular) off-center resolution. The resolution in the iso-
center should be affected only by the widths of the source and detector elements.
Fig. 9.8 left shows MTF ρ
50%
measurements for modeling of only detector element
width (N
SW
= 1, N
DW
= 2), and modeling of both source and detector element
widths (N
SW
= 2, N
DW
= 2). Similarly to the cross-plane resolution case, the
resolution does not continue to improve after the ten first iterations. Interestingly,
in contrast to the cross-plane resolution case, higher spatial resolution is obtained
with B
S
than with P
T
. The highest increase in MTF ρ
50%
is approximately 10%,
and is obtained with modeling of both detector and source width (N
SW
= 2,
N
DW
= 2), and the standard backprojector B
S
. Fig. 9.8 shows MTFs after ten
iterations. Although the MTFs differ very little, one interesting observation is that
by modeling both source and detector element widths, the MTF for P
T
follows
the MTF for B
S
for low frequencies, but decays more rapidly for high frequencies.
The off-center resolution is affected not only by the source and detector widths,
but also by the gantry rotation (mainly azimuthal resolution) and source length
(mainly radial resolution). To see how the source length affects the result, consider
9.2 Multiple ray forward projection 121
0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
Distance to center (mm)
S
S
P
B
S
, N
SL
= 1
B
S
, N
SL
= 3
1 1.5 2 2.5 3
−0.05
0
0.05
0.1
Distance to center (mm)
S
S
P
P
T
, N
SL
= 1
P
T
, N
SL
= 3
Figure 9.7: SSPs for WLS-IFBP after ten iterations. Left: complete SSP. Right:
Blow-up of detail. Both B
S
and P
T
were used as backprojectors. In all plots,
the detector height is taken into account with N
DH
= 3. Except for N
SL
, other
multiple ray parameters equal one.
0 5 10 15
3.6
3.7
3.8
3.9
Iteration number
M
T
F
ρ
5
0
%
(
l
p
/
c
m
)
B
S
, N
SW
= 1
B
S
, N
SW
= 2
P
T
, N
SW
= 1
P
T
, N
SW
= 2
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
,
1
0
i
t
e
r
a
t
i
o
n
s
B
S
, N
SW
= 1
B
S
, N
SW
= 2
P
T
, N
SW
= 1
P
T
, N
SW
= 2
Figure 9.8: Left: iso-center MTF ρ
50%
values for WLS-IFBP. Both B
S
and P
T
were used as backprojectors. Right: MTFs after ten iterations. In all plots, the
detector width and height are taken into account with N
DH
= 3 and N
DW
= 2.
Except for N
SW
, other multiple ray parameters equal one.
122 Data acquisition modeling
0 5 10 15
2.6
2.8
3
3.2
3.4
3.6
Iteration number
M
T
F
ρ
5
0
%
(
l
p
/
c
m
)
B
S
, N
SW
= 1, N
SL
= 1
B
S
, N
SW
= 2, N
SL
= 3
P
T
, N
SW
= 1, N
SL
= 1
P
T
, N
SW
= 2, N
SL
= 3
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
,
1
0
i
t
e
r
a
t
i
o
n
s
B
S
, N
SW
= 1, N
SL
= 1
B
S
, N
SW
= 2, N
SL
= 3
P
T
, N
SW
= 1, N
SL
= 1
P
T
, N
SW
= 2, N
SL
= 3
Figure 9.9: Left: off-center radial MTF ρ
50%
values for WLS-IFBP. Both B
S
and P
T
were used as backprojectors. Right: MTFs after ten iterations. In all
plots, the detector element widths and heights are modeled with N
DW
= 2 and
N
DH
= 3. Gantry rotation should not affect the off-center radial resolution, and
has not been modeled here, i.e. N
GR
= 1.
an object located 150mm off-center. This object is irradiated at a maximum fan-
angle of β = sin
−1
(150mm/R
S
). Ignoring the width of the source, at this fan-angle
the object “sees” the source as if it is l
S
150mm/R
S
≈ 3mm wide. Clearly, the
width of the resulting x-ray beam will cause blurring in the radial direction.
The left of Fig. 9.9 shows how off-center radial MTF ρ
50%
values are affected
by modeling of the detector elements and source, as well as backprojector type.
Only modeling the spatial extent of the detector results in an increase in ρ
50%
of
approximately 3%, independent of the backprojector type. When also the length
and width of the source are taken into account, the improvements of ρ
50%
values
increase to 33% for B
S
and 22% for P
T
. As shown in the MTF plots to the right,
the improved modeling does not change the location of the first zero in the MTF.
We also note that with modeling of source width and height, the MTF drops off
faster with P
T
than with B
S
.
Fig. 9.10 shows off-center azimuthal MTF ρ
50%
values and MTFs with and
without gantry rotation modeling. Without gantry rotation modeling, the ρ
50%
values increase with 3.5% for P
T
and 5.2% for B
S
. With gantry modeling, corre-
sponding numbers are 6.0% and 10% respectively.
Due to the large number of rays involved, many of the models presented here
are computationally too expensive to be applied to any full size clinical data set.
Since modeling of the detector height and width gives an easily observable increase
9.2 Multiple ray forward projection 123
0 5 10 15
2.8
2.9
3
3.1
3.2
Iteration number
M
T
F
ρ
5
0
%
(
l
p
/
c
m
)
B
S
, N
GR
= 1
B
S
, N
GR
= 2
P
T
, N
GR
= 1
P
T
, N
GR
= 2
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Frequency (lp/cm)
M
T
F
,
1
0
i
t
e
r
a
t
i
o
n
s
B
S
, N
GR
= 1
B
S
, N
GR
= 2
P
T
, N
GR
= 1
P
T
, N
GR
= 2
Figure 9.10: Left: off-center azimuthal MTF ρ
50%
values for WLS-IFBP. Both
B
S
and P
T
were used as backprojectors. Right: MTFs after ten iterations. In all
plots, the detector element widths and heights are modeled with N
DW
= 2 and
N
DH
= 3. The source length and width are modeled with N
SL
= 3 and N
SW
= 2.
in cross-plane resolution at a reasonable cost increase, we choose this model for a
visual inspection of physical head phantom reconstructions. Fig. 9.11 shows recon-
structed images of the physical head phantom with linear regularization. We note
that the detector element modeling improves the cross-plane resolution (indicated
with solid circles). Unfortunately, another obvious and immediate consequence of
the detector element modeling is an increased amount of wind-mill artifacts.
Fig. 9.12 shows how nonlinear regularization can partially be used to control
the increased amount of wind-mill artifacts. The upper row, i.e. Fig. 9.12 a)
and b), shows axial images obtained with linear and non-linear regularization,
but without any modeling of the spatial extent of the detector elements. In the
lower row, the detector element geometry has been modeled with N
DW
= 2 and
N
DH
= 3. Fig. 9.12 c) shows the result obtained with B
S
, and Fig. 9.12 d) shows
the result obtained with P
T
. In the area indicated by the solid circles, we see
that the high contrast cross-plane resolution is further improved by the detector
element modeling. The P
T
result has a slightly higher cross-plane resolution,
but as indicated by the dashed circles, this image also contains more wind-mill
artifacts.
9.2.3 Discussions and conclusions
Clearly, accurate modeling of the acquisition process can improve image recon-
struction in terms of spatial resolution. The amount of improvement depends
124 Data acquisition modeling
a) Axial image, B
S
N
DH
= 1, N
DW
= 1
b) Axial image, P
T
N
DH
= 3, N
DW
= 2
Figure 9.11: WLS-IFBP f
10
reconstructions of the physical head phantom us-
ing linear regularization. Reconstruction parameters: Q = 0.7, β = 0.47, β
xy
=
1.0, β
z
= 1.4. Greyscale window (40 ±75)HU. In both cases cases, N
SW
= N
SL
=
N
GR
= 1.
on which parts of the acquisition system that are modeled, and on the kind of
backprojector being used.
One recurring observation in the previous section is that, when the ray density
is high in relation to the voxel density, i.e., when Δ
t
< Δ
x
or Δ
q
h/2 < Δ
z
, higher
resolution is obtained with B
S
than with P
T
as backprojector. Vice versa, if the
voxel density is higher than the ray density, i.e. Δ
x
< Δ
t
or Δ
z
< Δ
q
h/2, higher
resolution is obtained with P
T
than with B
S
. In our scanner configuration, the
in-plane ray density is relatively high (Δ
x
= 0.5mm, Δ
t
= 0.4mm) and the cross-
plane ray density is relatively low (Δ
z
= 0.5mm, Δ
q
h/2 = 1.2mm). Consequently,
P
T
gives higher cross-plane resolution than B
S
, and B
S
gives higher in-plane
resolution than P
T
.
The difference in spatial resolution arising from using B
S
and P
T
as backpro-
jectors can be understood by looking at Fig. 9.13 and studying the fixed point
f

= (BWP+βR)
−1
BWp, where B equals either B
S
or P
T
. First we note that
input data frequency components belonging to the null space N(BW) cannot be
reconstructed. Of special interest is the lowest frequency in N(BW), correspond-
ing to the first zero in the amplitude spectrum of the backprojection interpolation
function. If the voxel density is higher than the ray density as in Fig. 9.13 c) and
d), this frequency is higher for P
T
than for B
S
. Equivalently, if the ray density is
higher than the voxel density as in Fig. 9.13 a) and b), this frequency is higher for
9.2 Multiple ray forward projection 125
a) B
S
, p = 2, q = 2, c = 20, β = 0.47
N
DW
= 1, N
DH
= 1
b) B
S
, p = 2, q = 1.3, c = 20, β = 0.47
N
DW
= 1, N
DH
= 1
c) B
S
, p = 2, q = 1.3, c = 30, β = 0.8
N
DW
= 2, N
DH
= 3
d) P
T
, p = 2, q = 1.3, c = 30, β = 0.8
N
DW
= 2, N
DH
= 3
Figure 9.12: Axial images of WLS-IFBP f
10
reconstructions using linear and
nonlinear regularization. Greyscale window (40 ± 75)HU. In all cases, N
SW
=
N
SL
= N
GR
= 1.
126 Data acquisition modeling
−1 −0.5 0 0.5 1
0
0.2
0.4
0.6
0.8
1
1.2
Distance (mm)
B
S
P
T
a) Spatial domain, in-plane
−5 0 5
0
0.2
0.4
0.6
0.8
1
1.2
Frequency (1/mm)
B
S
P
T
b) Fourier domain, in-plane
−1.5 −1 −0.5 0 0.5 1 1.5
0
0.2
0.4
0.6
0.8
1
1.2
Distance (mm)
B
S
P
T
c) Spatial domain, cross-plane
−2 −1 0 1 2
−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Frequency (1/mm)
B
S
P
T
d) Fourier domain, cross-plane
Figure 9.13: Interpolation functions and their Fourier transforms for B
S
and
P
T
in the t-direction and q-direction respectively.
B
S
than for P
T
. Another reason is that the balance between the regularization
contribution and data contribution is affected by the amount of low-pass filtering
done by the interpolation in the backprojection. A backprojector that preserves
more high frequencies results in a sharper iterative reconstruction result.
Since the computational cost rapidly increases with the accuracy of the acqui-
sition model, only very simple multiple ray models can be used on full size clinical
data sets. Already detector element modeling with N
DW
= 2 and N
DH
= 3 re-
sults in a forward projection that is six times more expensive than with a single
ray model. Therefore, it does not seem reasonable to further improve the model
with source and gantry rotation modeling unless there are special requirements on
off-center resolution.
An interesting alternative to the multiple ray experiments made here is to
analytically derive interpolation schemes corresponding to geometrical models of
9.3 A strip integral projection method 127
the source, detector elements, and gantry rotation [14, 15]. This could result in
computationally cheaper ways to achieve accurate linear models of the acquisition
system. One problem with these methods is that it is difficult to take the exact
geometry into account, e.g. the tilted anode in Fig. 9.2. A very simplified variant
of this is kind of methods is found in the next section.
As previously discussed, the in-plane resolution is limited by initial rebinning
of input data. One possibility to further improve in-plane resolution could be
to entirely avoid rebinning by performing forward projection and backprojection
directly in the cone-beam geometry, and use a rampfilter derived for fan-beam
reconstruction. Since normalization would not be done correctly, the first iterates
would be severely corrupted by artifacts. An interesting future investigation would
be to examine how well these artifacts are corrected by the iterative loop, and how
the in-plane resolution would be affected.
9.3 A strip integral projection method
9.3.1 Method description
In this section, we present a forward projection method P that avoids DC-aliasing
in its corresponding backprojection P
T
by calculating strip integrals instead of
line integrals. The method was first presented in [104]. Similar methods for have
previously been presented by Hanson and Wecksung [33], Ohyama et al. [76],
Mueller et al. [68], Ziegler et al. [124, 125], and De Man and Basu [17].
For simplicity, the method is presented in the context of 2-D parallel beam
CT. Except for some small details, generalization to 3-D cone-beam or semi-
parallel geometry is straightforward. Let θ be the projection angle and let t
denote parallel displacement as described in Section 2.1. Then, the direction of
the rays is given by s(θ) =

−cos θ, −sin θ

T
and the displacement is given by
t(θ, t) = t

sin θ, −cos θ

T
. By putting s and t into (3.13), the footprint can be
expressed as
g(θ, t) = g(s(θ), t(θ, t)) =
1
Δ

R
b(t(θ, t) +ls(θ))dl. (9.8)
This footprint can be used to construct a line integrating projection operator by,
for each voxel, projecting its footprint onto the detector, and sample the projected
footprint in the points of the detector elements.
Hanson and Wecksung [33] showed how the footprint can be used for calculation
of strip integrals. By storing the integral I(θ, t) =

t
−∞
g(θ,
˜
t)d
˜
t, a strip integral
128 Data acquisition modeling
defined by the displacements t
1
and t
2
of the basis function b can be calculated as

t
2
t
1

R
b(t(θ, t) +ls(θ))dldt = Δ

t
2
t
1
g(θ, t)dt =
= Δ

t
2
−∞
g(θ, t)dt −

t
1
−∞
g(θ, t)dt

= Δ(I(θ, t
2
) −I(θ, t
1
)). (9.9)
The new projection operator, which we refer to as the strip integral method,
is based on the Joseph method [41] in the way that it employs the same footprint
function
g
J
(θ, t) =
1
ζ(θ)
Λ

t
Δ
x
ζ(θ)

, (9.10)
where ζ(θ) = max([ cos θ[, [ sin θ[), Λ(t) = (0, 1−[t[), and Δ
x
equals the reconstruc-
tion volume sampling distance. However, instead of sampling g
J
as in the original
Joseph method, g
J
is integrated along intervals corresponding to the detector sam-
pling distance Δ
t
according to the integration technique described above.
By combining the Joseph method with the strip integration technique, DC-
aliasing is avoided both in forward projection and backprojection. To see this,
we study the coefficients of the P-matrix in more detail. Let θ
i
and t
i
be the
parameters defining the ray corresponding to measurement i. Furthermore, let t
i,j
be the projection of pixel j onto the detector along the direction corresponding to
the ith measurement, i.e., t
i,j
= x
j
sin θ
i
− y
j
cos θ
i
. Noting that integrating the
footprint over an interval (t
i
−t
i,j
±Δ
t
/2) is equivalent to sampling the footprint
convolved with a rectangle function of width Δ
t
in the point t
i
−t
i,j
, we see that
the resulting coefficients of the P-matrix are given by
p
ij
=
Δ
x
Δ
t
(g
J,i
∗ Π
Δ
t
)
. .. .
=:h
i
(t
i
−t
i,j
). (9.11)
Here, g
J,i
(t) = g
J

i
, t) and Π
Δ
t
is a rectangle function of width Δ
t
. The function
h
i
describes how the pixels of the image contribute to the measurement i during
forward projection, and how the ith measurement contribute to the image during
backprojection. The Fourier transform of h
i
(t) is given by
H
i
(ρ) = Δ
2
x
sinc
2

x
ζ(θ
i
)ρ) sinc(Δ
t
ρ). (9.12)
In order to avoid DC-aliasing during backprojection, H
i
(ρ) must be zero for ρ =

−1
t
, k ∈ Z
+
`¦0¦, which is obviously accomplished by the sinc factor. DC-
aliasing during forward projection is avoided by the squared sinc factor as discussed
in Section 3.3.4.
Thanks to the integration technique presented in (9.9), the strip integral
method is well suited for an efficient computer implementation. Similarly to the
9.3 A strip integral projection method 129
Joseph Iootprint
Proiected Joseph
Iootprint

0

1

2

3
2

2

cos0

1

2
Figure 9.14: Illustration of how the image is divided in to lines ¦L
i
¦ along the
coordinate axis that intersect the ray direction with maximal angle. The projection
of the Joseph footprint onto these lines is always a triangle of length 2Δ
x
.
original Joseph method, the first step in the implementation is to divide the image
into a set of lines ¦L
i
¦ parallel to the coordinate axis maximally orthogonal to
the ray direction as shown in Fig. 9.14. Because of the stretching of the Joseph
footprint (see (9.10)), its projection onto the lines ¦L
i
¦ is always a triangle of
width 2Δ
x
. A brief pseudocode description of the algorithm is found in Listing
9.1.
Listing 9.1: Pseudocode for the strip integral method.
1 Calculate the integrated footprint table I
P
(t)
2 for al l pr oj e c t i on angl e s
3 Divide the image into lines along the coordinate axis maximizing its
4 smallest angle to the ray direction for all image lines
5 for al l image l i n e s
6 Determine the integration interval (t
1
, t
2
)
7 Determine which pixels that contribute to the ray
8 for al l c ont r i but i ng pi x e l s
9 Calculate the contribution from the current voxel as the
10 difference I
P
(θ, t
2
− x
j
) − I
P
(θ, t
1
− x
j
) (see Fig. 9.15)
11 end
12 end
13 end
130 Data acquisition modeling
Proiected and integrated Iootprint

1

¦

)

2
Figure 9.15: The contribution from the jth pixel to the indicated ray is given
by I
P
(θ, t
2
−x
j
) −I
P
(θ, t
1
−x
j
).
Table 9.1: Scanning and reconstruction parameters
Number of channels 336 & 672
Number of projections (1/2 turn) 290 & 580
Number of pixels 512 512
Detector width 500mm
Width of reconstructed volume 500mm
9.3.2 Experiments
For evaluation of the strip integral method, 2-D reconstructions of the Shepp-
Logan phantom [89] have been studied. Noise-free input data were generated from
a Shepp-Logan phantom with a length of 368mm. Each detector contribution was
calculated as a mean value of seven parallel sub-rays. Other parameters are listed
in Table 9.1.
The method used for reconstruction was a steepest descent method for mini-
mizing the cost function
z(f ) =
1
2
|Pf −p|
2
W

N
¸
i=1
N
¸
j=1
d
ij
(f
i
−f
j
)
2
2
. (9.13)
This is equivalent to the WLS-IFBP method from Chapter 7 with v
i
= 1, i =
1, ..., N, and with optimal choices of the step lengths α
k
as described in (7.15).
Four different forward projection/backprojection pairs were compared, namely
1) P: Joseph, B: Standard backprojector B
S
.
2) P: Joseph, B: Transpose P
T
.
3) P: Strip integral, B: Transpose P
T
4) P: Distance driven (see below), B: Transpose P
T
The distance driven projection method by De Man and Basu [17] is very similar to
the strip integral method presented here. The only difference is that the distance
9.3 A strip integral projection method 131
driven projection operator uses the rectangular footprint
g
DD
(θ, t) =
1
ζ(θ)
Π

t
Δ
x
ζ(θ)

, (9.14)
while the strip integral method uses the slightly more complicated triangular foot-
print g
J
given by (9.10).
Image quality was studied by visual inspection and measurement of σ
e
values
(see Section 4.2) over (i) the complete image and (ii) regions containing only low
contrast structures.
Fig. 9.16 shows reconstructed images of the Shepp-Logan after ten iterations.
The number of channels is 336 and the number of views per half turn is 290. This
implies a detector sampling distance Δ
t
= 500/336mm ≈ 1.49mm. In comparison,
the image sampling distance is Δ
x
= 500/512mm ≈ 0.98mm, i.e. Δ
x

t
≈ 0.66.
The lowest artifact level, measured by σ
e2
is obtained with the strip integral
method. Furthermore, the σ
e1
measurements, reflecting errors around edges, in-
dicate that the images in Fig. 9.16 c) and d) are sharper than those in Fig. 9.16
a) and b).
In Fig. 9.17, the number of channels has been increased from 336 to 672, and
the number of views has been increased from 290 to 580. Therefore, Δ
t
equals
0.75mm, and Δ
x

t
= 1.31. Also in this case, the σ
e2
value is lower for the
strip integral method than for the other methods. However, in contrast to Fig.
9.16, no improvement of spatial resolution can be observed for the distance driven
method or the strip integral method. Instead, the Joseph method combined with
the standard backprojector gives the lowest σ
e1
value.
9.3.3 Discussions and conclusions
From the results in the previous section, it is clear that the strip integral method
can improve reconstruction results by reducing artifacts due to DC-aliasing in
backprojection, or aliasing of false high frequencies in the basis function during
forward projection. Furthermore, the strip integral method can improve spatial
resolution by taking into account the blurring during acquisition in a better way
than the original Joseph method. Although not investigated here, the improve-
ment in resolution is expected to be accompanied by increased noise levels as
observed for the multiple ray reconstructions in Fig. 9.11. By combining the
strip integral method with the nonlinear regularization presented in Chapter 8,
amplification of noise can probably be avoided without loosing the improved high
contrast resolution.
So far, the strip integral method has only been evaluated on a parallel-beam
geometry. Since the width of the strip in the strip integral method adapts to
the local ray density, it is assumed to be especially useful for avoiding aliasing
associated with divergent ray geometries such as the semi-parallel geometry or
132 Data acquisition modeling
a) P : Joseph B : B
S
σ
e1
= 79HU σ
e2
= 1.4HU
b) P : Joseph B : P
T
σ
e1
= 78HU σ
e2
= 1.5HU
c) P : Dist. dr. B : P
T
σ
e1
= 76HU σ
e2
= 1.6HU
d) P : Strip int. B : P
T
σ
e1
= 76HU σ
e2
= 1.2HU
Figure 9.16: Reconstructed images of the Shepp-Logan phantom after ten itera-
tions. N
ch
= 336 and N
proj
= 290. Greyscale window (0HU, 40HU). σ
e1
measures
errors over the whole image, while σ
e2
measures errors on low contrast regions.
9.3 A strip integral projection method 133
a) P : Joseph B : B
S
σ
e1
= 55HU σ
e2
= 3.8HU
b) P : Joseph B : P
T
σ
e1
= 56HU σ
e2
= 3.3HU
c) P : Dist. dr. B : P
T
σ
e1
= 56HU σ
e2
= 4.8HU
d) P : Strip int. B : P
T
σ
e1
= 56HU σ
e2
= 1.7HU
Figure 9.17: Reconstructed images of the Shepp-Logan phantom after ten itera-
tions. N
ch
= 672 and N
proj
= 580. Greyscale window (0HU, 40HU). σ
e1
measures
errors over the whole image, while σ
e2
measures errors on low contrast regions.
134 Data acquisition modeling
the cone-beam geometry (see Fig. 3.6). Therefore, future research should include
investigations on divergent beam geometries.
Our experiments show that aliasing artifacts caused by the discontinuities in
the rectangular footprint g
DD
of the distance driven method can be reduced by
switching to the triangular footprint g
J
. This is explained by the fact that the
Fourier transform of g
J
contains an extra sinc factor suppressing frequencies above
the Nyquist frequency of the reconstructed images. More remarkable is the result
in Fig. 9.17, showing stronger aliasing artifacts for the distance driven projector
than the Joseph projector, since this seems inconsistent with the results presented
in the original publication [17]. One possible explanation could be that the orig-
inal publication consider a cone-beam geometry where the adaptive anti-aliasing
properties of the distance driven projection method should help improving the
reconstruction.
From a performance point of view, we expect the strip integral method to be
slower than the distance driven method, but faster and simpler than the more
accurate and more aliasing suppressing blob based methods by Mueller et al. [68]
and Ziegler et al. [124, 125]. Efficient hardware implementations of the strip
integral method have not yet been considered, but the fact that the method can
be expressed as a series of resampling operations accessing data in a local and
predictive way seems to make it a good candidate for efficient implementation on
parallel processing hardware such as modern graphics board.
One interesting property of the integration technique (9.9) is that it can be
repeated to efficiently calculate convolutions with several rectangle functions. For
example, assume that we want to calculate the convolution of some function f
with the two rectangles Π
Δ
1
and Π
Δ
2
. Then,
(f ∗ Π
Δ
1
∗ Π
Δ
2
)(t) =
=

t+Δ
2
t−Δ
2

t


1
t

−Δ
1
f(t

)dt

dt

=

t+Δ
2
t−Δ
2
f
1
(t

+ Δ
1
) − f
1
(t

− Δ
1
)dt

=
= f
2
(t + Δ
2
+ Δ
1
) − f
2
(t + Δ
2
− Δ
1
) − f
2
(t − Δ
2
+ Δ
1
) + f
2
(t − Δ
2
− Δ
1
),
(9.15)
where f
1
(t) =

t
−∞
f(t

)dt

and f
2
(t) =

t
−∞
f
1
(t

)dt

. Thus, having pre-calculated
a table of f
2
, for any choice of parameters Δ
1
and Δ
2
the function (f ∗Π
Δ
1
∗Π
Δ
2
)
can be evaluated by looking up four values in the f
2
table. This could enable
efficient implementation of advanced analytical forward projection methods such
as those presented by Danielsson and Sunneg˚ardh in [15], and is another possible
subject for future research.
Chapter 10
Implementation on Graphics
Processing Units (GPU)
10.1 Introduction
The computationally most expensive operations in iterative reconstruction are the
forward projections and the backprojections. Given that input and output data
dimensions are proportional to some number N, the number of non-zero entries
in the P and B matrices is O(N
4
). Clearly, even for small problems, the number
of computations will be huge. For instance, the number of bilinear interpolations
needed in the Joseph forward projector P and backprojector P
T
in the physical
head phantom experiments presented earlier in the thesis equals
N
ch
N
rows
N
proj−tot
N
x
= 736 32 4267 512 ≈ 5.1 10
10
. (10.1)
The corresponding number for the standard backprojector B
S
is approximately
N
x
N
y
N
z
N
proj−irrad
≈ 512 512 250 2300 ≈ 1.5 10
11
, (10.2)
where N
proj−irrad
is the approximate number of projection angles for which a voxel
is irradiated.
To reduce the number of unique coefficients to be stored, backprojection to a
polar grid has previously been suggested [34]. If the number of angles in the polar
grid equals the number of projection angles, the same interpolation and weighting
coefficients can be used for all axial images. A similar trick was suggested by
Steckmann et al. [99], who suggested backprojection to a stack of axial images
rotated with a z-dependent angle φ = φ
0
+ 2πz/P. These techniques are very
efficient. Furthermore, since all coefficients are pre-calculated, they also allow for
complex models of the acquisition process without disproportionate performance
penalties. The main drawbacks of these methods are that they require relatively
large amounts of memory for look-up tables, the backprojected images must be
136 Implementation on Graphics Processing Units (GPU)
resampled to a Cartesian grid, and look-up tables must be regenerated for any
new Region Of Interest (ROI).
Without creating and exploiting symmetries and partially abandoning the
Cartesian grid as described above, the coefficients of the B and P matrices can-
not be stored, and need to be calculated on the fly. To handle the computational
load, commercial manufacturers of CT systems use backprojectors implemented ei-
ther on Application Specific Integrated Circuits (ASIC) or on Field-Programmable
Gate Arrays (FPGA) [19, 79]. These implementations are very fast, but have the
drawbacks of being expensive, inflexible, and therefore not suitable for research.
Instead, most researchers are referred to perform experiments on desktop work-
stations or, in some cases, supercomputing clusters.
Running experiments on full size clinical data sets on desktop workstations can
be painfully slow, especially in the case of iterative reconstruction. For example,
one single WFBP backprojection of the physical head phantom data set takes
almost four hours on a desktop workstation with one Intel Xeon E5440 processor
1
.
Therefore, for experiments on realistic input data, it is important to find simple
and inexpensive solutions for reducing the computation time.
Several researchers have reported on successful attempts on reducing the com-
putation times by making use of new parallel processing architectures such as the
Cell Broadband Engine (CBE) [43, 3, 87] and Graphics Processing Units (GPUs)
[67, 116, 53, 86]. For modern GPUs, results indicate that a one can obtain speed-
ups of between one and two orders of magnitude compared to single core CPU
implementations.
In 2006, the company NVIDIA released the Compute Unified Device Architec-
ture (CUDA). CUDA is a set of tools for general purpose programming of GPUs.
Compared to other well-known libraries such as OpenGL and DirectX, the lan-
guage of CUDA is more similar to C, and requires less prior knowledge about
graphics programming. Therefore, we have chosen to base our implementations of
forward projection and backprojection on CUDA.
In the following, we present and discuss our experiences from implementing
the operators P, B
S
, and P
T
. Note that the implementations presented here are
by no means optimal. The purpose of this chapter is to discuss the possibilities
of accelerating different operations with CUDA, rather than to provide optimal
implementations.
10.2 The GPU from a CUDA perspective
To understand how a GPU can improve the performance of some computationally
intensive programs by factors ranging from ten to one hundred, we first take a
1
Our CPU backprojection implementation uses one processor core, and does not utilize the
vectorization capabilities of the processor. No serious attempts have been made to optimize the
implementation. A more carefully designed single core implementation could be expected to run
up to four times faster.
10.2 The GPU from a CUDA perspective 137
DAM
Cache
CPU
Control
ALU
ALU
ALU
ALU
GPU
DAM
Figure 10.1: Usage of chip area on the CPU and GPU. Redrawn from the CUDA
programming guide [75].
look at the design of the GPU, and compare it with the CPU. Much of the ma-
terial in this and the following section has been taken from the NVIDIA CUDA
programming guide [75].
Fig. 10.1 schematically illustrates how the processor chip area is used on a
CPU and a GPU respectively. Maybe the most noticeable difference is that the
GPU devotes more chip area for Arithmetic Logic Units (ALUs), while the CPU
uses a larger area for cache and control. As indicated in the figure, the ALUs
are grouped into several Streaming Multiprocessors (SMs), to be further discussed
in the next paragraph. Within each SM, the ALUs work in a Single Instruction
Multiple Data (SIMD) fashion. This property together with the large number of
ALUs, makes the GPU well suited to perform simple but arithmetically intensive
parallel computations on large amounts data. As an example, the maximum pos-
sible number of Floating Point Operations per Second (FLOP/s) for the NVIDIA
GT200 processor is close to one TFLOP/s, which is almost a factor ten higher
than the corresponding number for a modern CPU.
Each SM has eight scalar processors (SPs) controlled by one instruction unit
as shown in Fig. 10.2. All SPs run the same set of instructions, but on different
data. In order to fully utilize the hardware, it is therefore important to try to
avoid programs that often require the parallel processes (threads) to take different
paths.
The maximum rate at which data that can be read by the GPU from the
global GPU memory can be as high as 159Gb/s on modern GPUs. However,
there is a significant latency for random fetches. Therefore, it can take 400-600
clock cycles to obtain a requested data value from the global GPU memory. To
avoid performance degradation due to this latency, each SM is equipped with a
constant cache, a texture cache, and the so called shared memory as shown in
Fig. 10.2. The constant cache speeds up memory access from a 64kB read-only
region of the global GPU memory. The texture cache improves memory access
from 1-D, 2-D, and 3-D structures by prefetching data in the close vicinity of data
138 Implementation on Graphics Processing Units (GPU)
Device (GPU)
Device memory (GPU memory)
Streaming Multiprocessor N
Streaming Multiprocessor 2
Streaming Multiprocessor 1
Texture Cache
Constant Cache
Ìnstruction
Unit
Processor 1 Processor 2 Processor 8
Registers Registers Registers
Shared Memory
Figure 10.2: Hardware for CUDA. Redrawn from the CUDA programming guide
[75].
being read. Another interesting property of the texturing hardware is that it also
implements 1-D, 2-D, or 3-D linear interpolation, and simple and fast handling of
boundary conditions. The third possibility of speeding up memory accesses involve
the shared memory, which is a 16kb memory per SM, accessible for all SPs within
the SM for both reading and writing. To accelerate memory access with shared
memory, the programmer has to coordinate manually the loading of data needed
by the SPs into the shared memory. Although this memory can be very useful for
some applications, it is rather complicated to use, and will not be discussed in the
following sections.
10.3 Programming CUDA
10.3.1 Kernels and threads
In the context of CUDA, the CPU and the GPU are commonly referred to as the
host and the device respectively. The host executes the program responsible for
transferring data to and from the device, and calls programs to be executed on the
device. A program that is executed on the device is called a kernel. When a call
to a kernel is executed, several instances of the kernel are started simultaneously
on the device. Each such instance is called a thread.
Listing 10.1 shows a simple example of a kernel and how it is called from the
host. The first five lines defines a kernel that performs a pointwise addition of the
elements in the arrays A and B, and stores the result in C. All arrays are assumed to
have the length N, and to be stored on the device memory, i.e., the global memory
10.3 Programming CUDA 139
Listing 10.1: C/CUDA code for adding two vectors. The code is a slightly
modified version of CUDA programming guide [75].
1 __global__ void vecAddKernel(float
*
A, float
*
B, float
*
C)
2 {
3 int i=threadIdx.x;
4 C[i]=A[i]+B[i];
5 }
6
7 int main()
8 {
9 // Missing: code that sets up the device, allocates
10 // memory on the GPU and copies input data (A,B) from
11 // the host to the device.
12
13 // Kernel invocation
14 vecAddKernel<<<1, N>>>(A,B,C);
15
16 // Missing: code that copies the output data (C) from
17 // the device to the host.
18 }
of the GPU. To define the kernel, the function qualifier global is used. This
means that the following function should be run on the device, and callable from
the host only. When the kernel is invoked on line 14, N threads are created and
executed on one single SM. The built in variable threadIdx on line three tells
each thread on which data to work, much like an index variable in a for-loop.
10.3.2 Thread organization
To make use of several SMs, and to allow for vectors larger than 512 elements, the
threads have to be organized into several so called thread blocks. Fig. 10.3 shows
the basic organization of threads. Each time a kernel is invoked, a new grid of
thread blocks is instantiated. Only one grid, i.e. one kernel, can be run at a time
on the device. The dimensions of the grid and thread blocks are specified in the
call to the kernel by passing variables, e.g. gridDim and blockDim, within the
triple inequality signs as <<<gridDim, blockDim>>>. For example, to specify
the dimensions of Grid 1 in Fig. 10.3, gridDim should be set to (3, 2, 1) and
blockDim should be set to (4, 3, 1).
Listing 10.2 shows complete vector addition C/CUDA code that makes use of
all available SMs for large input vectors
2
. This is accomplished by dividing the
total number of threads into N/BLOCKSIZE| blocks of size BLOCKSIZE (lines 26
and 27). As the kernel is invoked, each of these blocks executes in parallel if there
are enough SMs. The remaining blocks are put in a queue, and are executed when
the SMs become available.
2
The number of elements in the vector must be greater than or equal to the thread block size
times the number of SMs on the device of interest.
140 Implementation on Graphics Processing Units (GPU)
Listing 10.2: Complete C/CUDA code for a function performing pointwise ad-
dition of two float arrays. In contrast to Listing 10.1, all available SMs are
used.
1 #define BLOCKSIZE 256
2
3 __global__ void vecAddKernel(float
*
Ad, float
*
Bd, float
*
Cd, int N)
4 {
5 int i= (blockIdx.x
*
BLOCKSIZE) + threadIdx.x;
6 if (i < N)
7 Cd[i]=Ad[i]+Bd[i];
8 }
9
10 void vecAdd(float
*
Ah, float
*
Bh, float
*
Ch, int N)
11 {
12 // Find and initialize the device with maximum GFLOPS
13 cudaSetDevice( cutGetMaxGflopsDeviceId() );
14
15 // Allocate and copy memory from host to device
16 float
*
Ad;
17 cudaMalloc((void
**
) &Ad, N
*
sizeof(float));
18 cudaMemcpy(Ad, Ah, N
*
sizeof(float), cudaMemcpyHostToDevice);
19 float
*
Bd;
20 cudaMalloc((void
**
) &Bd, N
*
sizeof(float));
21 cudaMemcpy(Bd, Bh, N
*
sizeof(float), cudaMemcpyHostToDevice);
22 float
*
Cd;
23 cudaMalloc((void
**
) &Cd, N
*
sizeof(float));
24
25 // Set up gridDim and blockDim
26 dim3 blockDim(BLOCKSIZE, 1, 1);
27 dim3 gridDim(N/BLOCKSIZE + (N\%BLOCKSIZE>0 ? 1:0), 1, 1);
28
29 // Kernel invocation
30 vecAddKernel<<<gridDim, blockDim>>>(Ad, Bd, Cd, N);
31 cutilSafeThreadSync();
32
33 // Copy the result back to the host
34 cudaMemcpy(Ch, Cd, N
*
sizeof(float), cudaMemcpyDeviceToHost);
35
36 // Free device memory
37 cudaFree(Ad);
38 cudaFree(Bd);
39 cudaFree(Cd);
40 }
10.3 Programming CUDA 141
Host Device
Grid 1
Block(0,0)
Grid 2
Kernel 1
Kernel 2
Block(1,0) Block(2,0)
Block(0,1) Block(1,1) Block(2,1)
Block(0,0) Block(1,0)
Block(0,1) Block(1,1)
Block(0,2) Block(1,2)
Block(1,1)
Thread
(0,0)
Thread
(1,0)
Thread
(2,0)
Thread
(3,0)
Thread
(0,1)
Thread
(1,1)
Thread
(2,1)
Thread
(3,1)
Thread
(0,2)
Thread
(1,2)
Thread
(2,2)
Thread
(3,2)
Figure 10.3: Execution of CUDA kernels, and organization of threads into blocks
and grids. Redrawn from the CUDA programming guide (2007) [74].
To handle vectors with lengths that are not multiples of the thread block size,
sufficiently many blocks are generated to generate an output data with a length
that is equal to the input data length rounded up to the nearest multiple of the
thread block size. The - if-statement on line 6 makes sure that no threads read
or write out of bounds. To investigate the effect of this conditional statement on
blocks for which all threads satisfy the condition, we iterated the kernel invocation
10000 times with N equal to 256 32768. Both with and without the conditional
statement, the total execution time was 15.40s, suggesting that the cost of the
if-statement is much smaller than the costs of creating threads, calculating the
index variable, and compute the result. For thread blocks containing a mixture
of threads that do and do not satisfy the condition, the solution in Listing 10.2
may have a negative effect on performance. However, in most cases the number of
threads not satisfying the condition are relatively few. Therefore, these negative
effects on performance are expected to be small.
142 Implementation on Graphics Processing Units (GPU)
In Listing 10.2, the thread block size equals 256, which is much larger than the
number of SPs per SM. This excessive amount of simultaneous threads per SM is
used by CUDA to hide memory access latencies. The device has a mechanism for
rapidly switching between one set of threads and another. Threads that are not
active have all data stored in registers, which amount to 8192 per SM on the G80
chips and 16384 on GT200 chips. Therefore, the number of threads per block must
be smaller than the total number of registers on the SM divided by the number
of registers used by one thread. One way to improve performance of a CUDA
program is therefore to minimize the number of required registers per thread, thus
making it possible to use more threads simultaneously on one SM.
10.3.3 Textures
Since textures are an important part of the implementations in the following sec-
tions, we briefly explain them here. In CUDA, a texture is a chunk of memory that
is connected to the texture hardware via a texture reference. The texture hardware
improves the performance of read access by caching, and provides simple handling
of boundary conditions as well as linear interpolation. Texture references can be
bound to 1-D, 2-D, and 3-D arrays. To get most out of the texture functionality,
the memory connected to the texture should be stored in a special structure called
cudaArray. A cudaArray can be accessed by copy commands and texture fetch
commands, but it cannot be directly read from or written to.
When a cudaArray is created in device memory and a texture reference is
bound to it, the type and dimensionality of data are specified, as well as the
interpolation to use and how boundaries should be handled. The interpolation
can be set to nearest neighbor or linear interpolation, and the boundaries can be
handled either by constantly extrapolating the boundary pixels of the texture,
or by extrapolation with repeated copies of the texture. In the kernel, a value
from the texture is obtained by using a texture fetch function (tex1D, tex2D, or
tex3D), taking the texture reference and a coordinate vector.
According to the CUDA programming guide, textures are optimized for 2-D
locality, meaning that improved performance is obtained if threads in the same
block or neighboring blocks read from neighboring locations in the texture. In re-
sampling operations, neighborhood operations, as well as the operations discussed
in the following sections, this condition is easy to meet.
10.3.4 Constant memory
As mentioned in Section 10.2, each SM provide a so called constant cache. More
specifically, the device has a total constant memory of 64kB. Each SM has a cache
of 8kB. On a cache hit, access to the constant memory is as fast as access to
a register. Therefore, this memory can be used for certain relatively small pre-
calculated lookup tables.
10.4 Forward projection 143
10.4 Forward projection
Implementation of the Joseph forward projection in CUDA is straightforward. We
begin explaining the program to be run on the host. Pseudocode for this program
is listed in Listing 10.3. For simplicity, it is assumed that there is enough space
on the host and device to store the whole voxel volume, and that only one line
integral per measurement is calculated.
Listing 10.3: Pseudocode for the forward projection. Host part.
1 Copy the input volume to the device, and bind a 3D-texture to it.
2 Configure grid and thread blocks.
3 Calculate constant memory lookup tables for simplified computation
4 of the x-ray source position (α,z).
5 for al l pr oj e c t i on angl e s
6 Calculate source position for the central ray (α
C
, z
C
).
7 Call the kernel which computes a projection image for the current angle.
8 Copy the projection image from the device and save it.
9 end
The first step in the host code is to allocate the required memory on the host
and the device, and to load the input volume. On the device, the input volume
is stored as a three-dimensional cudaArray, and bound to a texture reference.
This allows us to use the caching mechanisms of the texture hardware, as well
as the hardware linear interpolation. Since only 2D interpolation is required by
the Joseph method, it is not necessary to use a 3D texture. However, using a
3D texture makes the implementation simpler by avoiding use of several texture
references, and by avoiding the division of the implementation into two cases
depending on the direction of interpolation.
The second step is to set the gridDim and blockDim parameters that specify
the thread organization described in Section 10.3.2. We let one thread correspond
to one output value, i.e. one ray. The threads are grouped into blocks of 16
detector channels times 2 detector rows, and one grid represents one projection
angle.
The last preparational step of the host program is to calculate lookup tables
that simplify calculations in the kernel. For each ray/thread, the corresponding
source position has to be calculated. Given parallel projection angle θ, the α-angle
(see Fig. 2.6) of a ray with channel index chIx can be calculated as
α = θ −arcsin(t/R
S
) = θ −arcsin((t
0
+chIx Δ
t
)/R
S
), (10.3)
where t
0
is the t-value corresponding to chIx=0, and Δ
t
is the sampling distance
along the t direction. To reduce the number of registers required by the kernel, the
rightmost term can be pre-calculated and made accessible to the kernel through
144 Implementation on Graphics Processing Units (GPU)
the constant memory. In the same way, a lookup table for z = z
C
+(α−θ)P/(2π)
can be pre-calculated, where P is the patient table feed per turn, or equivalently,
the pitch of the helix.
When all preparations are finished, the main loop over all projection angles
starts. For each angle, the θ angle and the z for the x-ray source corresponding to
the central ray are calculated. In the kernel invocation, these variables are sent as
parameters to tell the kernel which projection image that shall be calculated. As
the calculation is finished, the result is copied to the host memory and saved to
file.
Pseudocode for the kernel is listed in Listing 10.4. As previously mentioned,
the task of one instance of this program is to calculate one line integral. The
line of interest is specified by the built in variables blockIdx and threadIdx
variables, which in an initial step are transformed into a channel index chIx and
a row index rowIx. The rest of the kernel program does not differ much from
the inner loop in a normal CPU implementation, except from the interpolation
step on line 10. In CUDA, this reduces to a call to the texture hardware, which
performs the actual interpolation.
Listing 10.4: Pseudocode for the forward projection. Device/kernel part. This
code correspond to the calculation of one line integral.
1 Calculate channel and row indices (chIx,rowIx) from blockIdx and threadIdx.
2 Calculate source position with help from the pre-calculated constant memory tables.
3 Calculate a normalized ray direction vector
4 Transform the source position and normalized ray direction to the coordinate
5 system of the voxel volume.
6 Calculate the first and last intersection between the ray and the contributing
7 image planes.
8 for al l c ont r i but i ng image pl anes
9 Calculate intersection between the ray and the contributing plane.
10 Fetch a contribution from the 3D texture bound to the input volume.
11 Accumulate the contribution.
12 end
13 Perform length scaling.
14 Save measurement to the device memory.
To examine the performance of the implementation, we generated projection
data from the clock phantom and the head phantom from previous chapters, on a
system with an Intel Xeon E5440 processor, and an NVIDIA Tesla C870 computing
card. The computation times for the clock phantom were 5800s for the CPU
implementation and 81s for the GPU implementation. Corresponding numbers for
the head phantom were 2700s and 32s respectively. In other words, performance
improvements of a factor of 72 and a factor of 84 were observed.
10.5 Backprojection 145
Listing 10.5: Pseudocode for the WFBP/standard backprojector. Host part.
1 Copy the input projection data to the device, and bind a 3D-texture to it.
2 Configure grid and thread blocks.
3 Calculate constant memory and texture lookup tables for simplified
4 computations of voxel projections onto the virtual detector (Fig. 10.4).
5 for al l a xi a l i mages
6 Calculate the z-coordinate of the image.
7 Call the kernel which computes the axial image.
8 Copy the projection image from the device and save it.
9 end
10.5 Backprojection
10.5.1 The WFBP/standard backprojector
In many ways, the CUDA implementation of the WFBP and standard backpro-
jector is similar to the CUDA implementation of the Joseph forward projector.
Pseudocode for the host part is listed in Listing 10.5.
As for the forward projection implementation, the preparations consist of load-
ing input data to the device and bind it to a 3D texture, set up the parallelization
(gridDim,blockDim), and calculating lookup tables. In our WFBP/standard
backprojection implementation, one thread corresponds to one voxel, one thread
block corresponds to 8 8 neighboring threads in an axial image, and one grid
consists of all threads needed to create one axial image.
Fig. 10.4 shows the backprojection geometry. To simplify the calculation of
the projection of a voxel onto the virtual detector, both constant memory lookup
tables and texture lookup tables are used. In the kernel, the calculation of the
channel index chIx is done according to
chIx =
x
0
sin θ −y
0
cos θ −t
0
Δ
t
. .. .
T
ch,1
(θ)
+xIx
Δ
x
Δ
t
sin θ
. .. .
T
ch,2
(θ)
+yIx
−Δ
x
Δ
t
cos θ
. .. .
T
ch,3
(θ)
, (10.4)
where T
ch,1
, T
ch,2
, and T
ch,3
indicate pre-calculated constant memory tables that
depend on the variable θ. Note that x and xIx are related by the equation
x = x
0
+xIx Δ
x
. Equivalent relationships apply to y, t, and q.
The detector row index rowIx is calculated as
rowIx =
R
S
Δ
q
m
(z −z
S
) −
q
0
Δ
q
=
=
1
Δ
q
cos β
. .. .
T
row,1
(t)

Δ
q
l
R
S

z
0
+zIx Δ
z

z
S
0

βP

....
T
row,2
(t)


q
0
Δ
q
, (10.5)
146 Implementation on Graphics Processing Units (GPU)

z
R
S
¦ x
S
. v
S
. z
S
¦ x
S
. v
S
. z
S

¦ x . v . z
l
m
h
q
¦ x . v . z

t

max

Figure 10.4: The backprojection geometry.
where T
row,1
and T
row,2
are t-dependent tables that are stored in the texture mem-
ory. The term
Δ
q
l
R
S
is given by
Δ
q
l
R
S
=
Δ
q
R
S
(x
0
cos θ +y
0
sin θ)
. .. .
T
row,3
(θ)
+xIx
Δ
q
Δ
x
R
S
cos θ
. .. .
T
row,4
(θ)
+yIx
Δ
q
Δ
x
R
S
sin θ
. .. .
T
row,5
(θ)
. (10.6)
Similarly to the channel index calculation, T
row,3
, T
row,4
, and T
row,5
represent
θ-dependent constant memory tables.
The main loop reconstructs one 2D-image at a time. For each image, the corre-
sponding z-value is calculated, and passed as an argument to the kernel invocation.
When the device code has produced an image, it is copied back to the host and
saved to file.
Listing 10.6 shows pseudocode for the kernel part of the WFBP/standard back-
projector. Lines marked WFBP perform the weighting and normalization that are
specific to the WFBP method.
The first step in the kernel is to calculate the voxel indices xIx and yIx from
blockIdx and threadIdx. This is followed by the main loop running over all
projection angles
˜
θ from 0 to π. For each projection angle
˜
θ, the inner loop covers
all source positions corresponding to this angle, and the angle
˜
θ + π. For each
source position, the voxel is projected onto the virtual detector, i.e., the channel
index chIx and row index rowIx are calculated. If the projection of the voxel
lies within the boundaries of the virtual detector, the detector value is read and
accumulated.
To examine the performance of the backprojector, the clock and head phantoms
were reconstructed, using the scanning geometry described in Chapter 4. The
10.5 Backprojection 147
Listing 10.6: Pseudocode for the WFBP/standard backprojector. Device/kernel
part.
1 Calculate xIx and yIx from blockIdx and threadIdx.
2 for
˜
θ = 0

to 180

3 Calculate chIx for θ =
˜
θ (see (10.4)).
4 for θ =
˜
θ,
˜
θ + 360

,
˜
θ + 720

, ...
5 Calculate rowIx (see (10.5) and (10.6)).
6 i f rowIx>0 and rowIx<Nrows−1
7 Fetch a contribution from the 3D texture bound to input data
8 WFBP: apply Q-weight
9 WFBP: accumulate Q-weight
10 end
11 end
12 Calculate chIx for θ =
˜
θ + 180

(see (10.4)).
13 for θ =
˜
θ + 180

,
˜
θ + 540

,
˜
θ + 900

, ...
14 Calculate rowIx (see (10.5) and (10.6)).
15 i f rowIx>0 and rowIx<Nrows−1
16 Fetch a contribution from the 3D texture bound to input data
17 WFBP: apply Q-weight
18 WFBP: accumulate Q-weight
19 end
20 end
21 WFBP: Divide the pixel value with the accumulated Q-weight
22 end
computation times were 25000s and 15000s for the CPU implementation, and 96s
and 84s for the GPU implementations. This corresponds to speedups of 260 and
180 times respectively.
10.5.2 The Joseph backprojector P
T
To understand what the Joseph backprojector does, we begin considering a row
in the matrix of the Joseph forward projection P. This row holds the coefficients
describing how the different voxels in the voxel volume contribute to one projection
data element during forward projection. When P is transposed, this row turns
into a column in P
T
. This column defines how P
T
maps one projection data
element into the voxel volume. In other words, during backprojection with P
T
,
one projection data element contributes to the voxel volume with the exact same
coefficients, as the voxel volume contribute to this projection data element during
forward projection with P.
P
T
seems to be more difficult to implement efficiently in CUDA than P and
B
S
. First, consider the possibility to let one output data element, i.e. one voxel,
correspond to one thread as in the previous section. This approach is difficult
because the task of finding the rays contributing to the output voxel becomes very
148 Implementation on Graphics Processing Units (GPU)
Joseph Iorward proiection P Joseph backproiection P
T
Figure 10.5: The difference between Joseph forward projection and backprojec-
tion.
Block 0 Block 1 Block 2

Figure 10.6: Division of one projection image into blocks to be executed in a
sequence.
complicated even for simple ray geometries. The second possibility would be to
reuse the procedure of the forward projection, i.e. let one thread correspond to
one ray, but replace the reading commands with writing commands (see Fig. 10.5).
This would be a simple and good approach for a single core CPU implementation,
but does not work for parallel implementations since there would be conflicts
between different threads wanting to update the same voxel.
To solve this problem, we divide the rays of one projection angle into blocks to
be executed in a sequence as illustrated in Fig. 10.6. Within each block, the rays
are sufficiently spatially separated so that no voxel update conflicts occur, thus
allowing for parallel execution. Since the number of rays per block is much smaller
than the total number of rays, the thread block configuration must be altered. In
our implementation, one thread corresponds to one voxel update of one ray, one
thread block corresponds to the total update of one ray, and one grid corresponds
to one “sequential” block.
Unfortunately, the performance of the above described approach is poor. The
computation times for the clock phantom were 7200s for the CPU and 2700s for
the GPU. Corresponding times for the head phantom were 3200s for the CPU and
750s for the GPU. This corresponds to speedups of 2.7 times and 4.2 times.
10.6 Discussions and conclusions 149
10.6 Discussions and conclusions
The results from the previous sections show that the computational performance
of the Joseph forward projection and the WFBP/standard backprojection can be
improved significantly by using the GPU. For the forward projection, speedups
of around a factor of 80 were observed, and for the backprojection, speedups of
more than a factor of 180 were observed. Two important factors contributing to
this improved performance is (i) the hardware interpolation offered by the texture
hardware, and (ii) the caching provided by the texture hardware.
In the CUDA implementation of the transpose P
T
of the Joseph forward pro-
jection, we did not find any good use of the hardware interpolation or the texture
caches. Instead, each thread performs uncached read and write operations to four
different locations in the device memory with rather poor performance. Since
results from previous chapters have indicated that P
T
is useful in iterative recon-
struction, an interesting future research topic would be to search for more efficient
exact or approximate implementations of this operation.
When interpreting the speedup factors presented here, one must keep in mind
that both the CPU and GPU implementations are sub-optimal. The CPU imple-
mentations could be improved by using SSE extensions and utilizing more cores,
and it has been shown that the performance of GPU implementations may vary
much with the configuration of parallelization [84]. Nevertheless, it is clear that
graphics hardware can be used to accelerate time consuming operations in recon-
struction considerably.
150 Implementation on Graphics Processing Units (GPU)
Chapter 11
Conclusions and future
research
11.1 Summary of conclusions
In this thesis, we have developed and examined three IFBP methods for it-
erative improvement of image quality in helical cone-beam CT: the RIWFBP
method based on iterative application of the WFBP method, the OS-IWFBP, and
the WLS-IFBP method based on gradient descent minimization of a spectrally
weighted quadratic cost function. All three methods suppress cone artifacts and
alleviate windmill artifacts, but they can also alter spatial resolution and noise
properties. To avoid such unwanted effects, and to improve the convergence of the
methods, linear and non-linear regularization have been investigated. Further-
more, we have examined possibilities to further improve image quality by accurate
modeling of the projection operator.
The main conclusion of Chapter 5 is that for cone angles up to ±2.78

, cone ar-
tifacts can be perfectly suppressed within three iterations of the RIWFBP method.
Thanks to better handling of redundant data, the amplitude of windmill artifacts
is also reduced. Without regularization, each iteration increases the amount of
high frequencies and noise in the reconstructed images, but with linear regular-
ization, i.e. linear smoothing of the reconstruction result in each iteration, these
effects can be alleviated or completely avoided. Another positive effect of the reg-
ularization is the improved convergence properties, evident from comparing the
rate of change, measured as log
10
(|f
k+1
− f
k
|/C) in Fig. 5.15. In the course of
events we found that even with regularization, the RIWFBP reconstruction re-
sults suffer from overshoots in the z-direction. Fortunately, these can be avoided
by preprocessing of projection data with a low-pass filter in the q-direction.
To reduce the number of required iterations, the ordered subsets technique [39]
can be applied to RIWFBP, resulting in the OS-IWFBP method. In Chapter 6, it
152 Conclusions and future research
is shown that this technique can improve the rate of cone artifact suppression for
high cone angles (4.8

and 9.6

). However, in its current form, the OS-IWFBP is
non-regularized, and its convergence properties are uncertain.
Due to the non-symmetric QP matrix, the RIWFBP method is difficult to
formulate in terms of an optimization problem, and hence difficult to analyze,
to improve, and to compare with other iterative methods for tomographic recon-
struction. In Chapter 7, we present the WLS-IFBP method, which is similar to
the RIWFBP, but has a simplified normalization procedure. Thanks to the latter
feature, WLS-IFBP can be expressed as a gradient descent method for solving a
weighted least squares problem. Thereby not only improvements such as auto-
matic determination of the step lengths α
k
, but also more advanced algorithms
such as the conjugate gradient method, can be introduced. Unfortunately, certain
artifacts are slightly more pronounced for WLS-IFBP than for RIWFBP. However,
the difference is very small, and they could possibly be compensated for by a more
sophisticated weighting.
Nonlinear regularization can be used to obtain contrast dependent spatial res-
olution. In Chapter 8, it is described how to use nonlinear regularization with the
RIWFBP and WLS-IFBP methods. Experiments with WLS-IFBP show that the
spatial resolution for high contrast structures can be improved while preserving the
spatial resolution and signal-to-noise ratio for low contrast structures. Because of
the nonlinear regularization, noise with an amplitude that is lower than a certain
threshold is suppressed, while noise with an amplitude that is higher than this
threshold is amplified. Therefore, the regularization parameter must be chosen
with respect to the expected noise level. One problem with nonlinear regulariza-
tion is that more iterations are needed for convergence, and that the true optimum
might not be found due to the fixed step length α. However, despite this fact, a
considerable improvement is observed during the first ten iterations.
In the first part of Chapter 9, it is shown that accurate modeling of the acquisi-
tion process can improve image reconstruction in terms of spatial resolution. The
actual amount of improvement depends on the scanner geometry and on the type
of backprojector used. An interesting observation is that by modeling detector
element size and gantry rotation in P, lowpass filtering adapted to the projection
data sampling is introduced. This lowpass filtering helps to avoid DC-aliasing
during backprojection with P
T
. Therefore, backprojection with P
T
produces bet-
ter results for more accurate data acquisition models. Furthermore, P
T
has the
advantage of producing exactly the structures P can detect.
In the second part of Chapter 9, we present a projection operator based on strip
integrals. The main conclusion is that the strip integration, i.e. convolution with a
rectangle adjusted to the ray density (detector sampling), suppresses DC-aliasing
in both forward projection and backprojection, and therefore reduces aliasing ar-
tifacts in comparison with line-integration based projector/backprojector pairs.
Furthermore, since the strip integral to some degree takes into account the blurring
11.2 Suggestions for future research 153
that occurs during data acquisition, the strip-integration projector/backprojector
pair also improves spatial resolution.
The computationally most expensive part of a reconstruction pipeline is the
projection and backprojection operations. Our experiments on CUDA implemen-
tation of these operations in Chapter 10 have shown that the Joseph projector and
the standard backprojector can be accelerated with a factor of more than 80 com-
pared to a single core CPU implementation. However, the transpose P
T
of the
Joseph operator is more complicated to accelerate with CUDA. Unfortunately,
we have not managed to make use of texture caches or hardware interpolation.
Hence the gain in performance for implementation of P
T
stayed at the not overly
impressive factor of four.
11.2 Suggestions for future research
Conjugate gradient optimization. Unfortunately, some of the methods for
image quality improvement proposed in this thesis increase the number of required
iterations. In the case of nonlinear regularization, it may even happen that the
optimum cannot be reached by using a constant step length α. We believe that
both the final convergence and the rate of convergence may be improved by a more
advanced update method, such as the Conjugate Gradient (CG) method with line
search by Fessler and Booth [25]. We mention this particular method because,
similarly to the methods studied in this thesis, it has the attractive property of
requiring only one calculation of P and P
T
per full iteration. Thanks to the line
search, such a method would also eliminate the need for choosing a suitable α
depending on the regularization parameter β.
Spatially variable step lengths. It is not obvious that a step length α that
is constant for the whole reconstruction volume gives the highest rate of con-
vergence. For instance, preliminary experiments have shown that the number
of required iterations can be reduced by using higher step lengths close to high
contrast structures. Together with a line search method, this could improve the
convergence rate without risking divergence.
Improved weighting for WLS-IFBP. As shown in Chapter 7, the WLS-IFBP
generates slightly more pronounced artifacts than the RIWFBP method in some
cases. One possible cause for this could be that the normalization in the RIWFBP
method takes better care of the actual data redundancies. If this presumption
could be proved correct, the WLS-IFBP could be improved by replacing the current
Q-weighting with a more advanced weighting that takes data redundancies into
account.
Statistical weighting. All methods investigated in this thesis falsely assume that
the variance of a detector measurement is independent of its expectation. Many
authors have shown that by taking into account the dependence between variance
154 Conclusions and future research
and expectation, an improved signal to noise ratio can be obtained [109, 126].
Unfortunately, this comes to the price of a more complicated numerical problem.
An interesting topic for future research is to examine computationally inexpensive
possibilities to improve the noise models of IFBP methods.
Implementation friendly projection/backprojection pairs. In Chapter 9,
it was shown that using the adjoint matrix B = P
T
for backprojection sometimes
results in better spatial resolution. As previously mentioned, it also has the ad-
vantage of producing exactly the structures P can detect, and ensuring that BP
is positive definite. Therefore, it is of interest to find good projector/backpro-
jector pairs P/P
T
such that both P and P
T
can be efficiently implemented on
existing hardware. One example of such a projector/backprojector pair is pre-
sented in 9.3. Although we have not found an efficient exact implementation of
this projection/backprojection pair, a good approximation of both projector and
backprojector can easily be implemented in CUDA.
Native cone-beam geometry. Since rebinning to semi-parallel data is the first
step in the methods studied here, in-plane resolution is inevitably lost. By formu-
lating the WLS-IFBP method directly in the cone-beam geometry, and thereby
entirely avoid rebinning, there is hope to improve spatial resolution and suppress
in-plane aliasing artifacts.
Appendices
A Acronyms
Adaptive Multiple Plane Reconstruction AMPR p. 67
Algebraic Reconstruction Technique ART p. 21
Application Specific Integrated Circuit ASIC p. 136
Arithmetic Logic Unit ALU p. 137
Cell Broadband Engine CBE p. 136
Center for Medical Image Science and Visualization CMIV p. 2
Compute Unified Device Architecture CUDA p. 7
Computed Tomography CT p. 1
Conjugate Gradient CG p. 153
Contrast Recovery Coefficient CRC p. 22
Direct Fourier Method DFM p. 11
Electrical and Musical Industries limited EMI p. 1
Field-Programmable Gate Array FPGA p. 136
Filtered Backprojection FBP p. 3
Floating Point Operation FLOP p. 137
Full Width at Half Maximum FWHM p. 59
Graphics Processing Unit GPU p. 7
Hounsfield Units HU p. 4
Iterative Coordinate Descent ICD p. 67
Iterative Filtered Backprojection IFBP p. 4
Local Region Of Interest LROI p. 49
Markov Random Field MRF p. 34
Maximum a Posteriori MAP p. 22
Maximum Likelihood ML p. 22
Maximum Likelihood Expectation Maximization MLEM p. 22
156 Appendices
Modulation Transfer Function MTF p. 40
Ordered Subsets OS p. 6
Ordered Subsets Convex OSC p. 22
Ordered Subsets IWFBP OS-IWFBP p. 6
Point Spread Function PSF p. 11
q-Generalized Gaussian Markov Random Field qGGMRF p. 7
Region Of Interest ROI p. 50
Regularized Iterative Weighted Filtered Backprojection RIWFBP p. 45
Root Mean Squared Error RMSE p. 37
Scalar Processor SP p. 137
Signal to Noise Ratio SNR p. 20
Simultaneous Algebraic Reconstruction Technique SART p. 21
Simultaneous Iterative Reconstruction Technique SIRT p. 21
Single Photon Emission Computed Tomography SPECT p. 22
Slice Sensitivity Profile SSP p. 40
Streaming Multiprocessor SM p. 137
Weighted Filtered Backprojection WFBP p. 4
Weighted Least Squares IFBP WLS-IFBP p. 7
B Notation
Below follows a list of symbols and acronyms used in the thesis. Unfortunately
the symbols α and β mean both fan-beam projection coordinates and parameters
in the RIWFBP method. However, their meaning at a certain location should be
clear from the context.
Scalars: geometry and sampling
Reconstruction space coordinates x, y, z p. 3
Reconstruction space coordinate vector r p. 23
Parallel-beam projection data coordinates θ, t p. 9
Fan-beam projection data coordinates α, β p. 14
Projection data row coordinate q p. 15
Frequency coordinates u, v p. 10
Radial frequency coordinate ρ p. 10
Source-isocenter distance R
S
p. 2
Detector height h p. 3
Table feed per turn/helical pitch P p. 3
Maximal fan-angle β
max
p. 36
Maximal cone-angle κ
max
p. 3
Voxel volume dimensions (N
y
= N
x
) N
x
, N
z
p. 30, 135
Sample distances for voxel volume (Δ
y
= Δ
x
) Δ
x
, Δ
z
p. 23, 27
B Notation 157
Detector dimensions N
ch
, N
rows
p. 36
Sample distances, non-rebinned proj. data Δ
β
, Δ
q
p. 115
Sample distances, rebinned proj. data Δ
t
, Δ
q
p. 12, 115
Scalars: evaluation
Error measurement (RMSE) σ
e
p. 38
Noise measurement σ
n
p. 38
Set for error and noise measurements Ω
1
, Ω
2
p. 38
MTF 50% frequency ρ
50%
p. 57
Scalars: reconstruction parameters
Detector row down-weighting parameter Q p. 17
Regularization parameter β p. 34
Regularization parameter xy-plane β
xy
p. 48
Regularization parameter z-direction β
z
p. 48
Scalars: others
Step lengths α, α
k
p. 6, 19
Scalar valued functions
Projection data (2D) p(t, θ) p. 9
Basis function b(r) p. 23
Irradiation function w
i
(r) p. 24
Function to be reconstructed f(x, y, z), f(r) p. 6
Potential function V () p. 34
Vectors and matrices
Reconstruction matrix Q p. 6
Projection matrix P p. 6
Input projection data vector p
in
p. 6
Sequence of reconstruction results f
1
, f
2
, ... p. 5
Limit point for reconstruction results f

p. 48
Regularization matrix/operator R, R() p. 47
Transforms
1D Fourier transform T p. 10
1D Fourier transform (radial direction) T
t
p. 10
2D Fourier transform T
2
p. 10
Radon transform { p. 9
158 Appendices
Bibliography
[1] A. H. Andersen and A. C. Kak. Simultaneous algebraic reconstruction tech-
nique (SART): a superior implementation of the ART algorithm. Ultrasonic
Imaging, 6(1):81–94, 1984.
[2] F. J. Beekman and C. Kamphuis. Ordered subset reconstruction for x-ray
CT. Physics in Medicine and Biology, 46(7):1835–1844, 2001.
[3] T. M. Benson, W. Dixon, and S. Basu. 3D filtered backprojection for curved
detector axial cone-beam data on a Playstation 3. In IEEE Nuclear Science
Symposium Conference Record, volume 5, pages 3522–3525, 2007.
[4] A. Borsdorf, R. Raupach, T. Flohr, and J. Hornegger. Wavelet Based Noise
Reduction in CT-Images Using Correlation Analysis. IEEE Transactions on
Medical Imaging, 27(12):1685–1703, 2008.
[5] R. N. Bracewell. The Fourier transform and its applications. McGraw-Hill,
Singapore, 1986.
[6] R.N. Bracewell and A.C. Riddle. Inversion of fan-beam scans in radio as-
tronomy. Astrophysical Journal, 150:427–434, 1967.
[7] Y. Censor, T. Elfving, and G. T. Herman. A method of iterative data refine-
ment and its applications. Mathematical Methods in the Applied Sciences,
7(1):108–123, 1985.
[8] L. Chang. A method for attenuation correction in radionuclide computed
tomography. IEEE Transactions on Nuclear Science, 25(1):638–642, 1978.
[9] G.-H. Chen, J. Tang, and S. Leng. Prior image constrained compressed
sensing (PICCS): a method to accurately reconstruct dynamic CT images
from highly undersampled projection data sets. Medical Physics, 35(2):660–
663, 2008.
[10] P.-E. Danielsson, P. Edholm, J. Eriksson, and M. Magnusson. Towards exact
3D-reconstruction for helical cone-beam scanning of long objects. A new
detector arrangement and a new completeness condition. In 4th International
Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and
Nuclear Medicine. Pittsburgh, Pennsylvania, USA, 1997.
160 Bibliography
[11] P.-E. Danielsson, P. Edholm, J. Eriksson, M. Magnusson, and H. Turbell.
The original PI-method for helical cone-beam CT. Technical Report
LiTH-ISY-R-2085, ISSN 1400-3902, Department of Electrical Engineering,
Link¨ oping University, December 1998.
[12] P.-E. Danielsson and M. Magnusson. A proposal for combining FBP and
ART in CT-reconstruction. In 7th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. St
Malo, France, 2003.
[13] P.-E. Danielsson and M. Magnusson. Combining Fourier and iterative meth-
ods in computer tomography. Analysis of an iteration scheme. The 2D-case.
Technical Report LiTH-ISY-R-2634, ISSN 1400-3902, Department of Elec-
trical Engineering, Link¨ oping University, March 2005.
[14] P.-E. Danielsson, M. Magnusson, and J. Sunneg˚ardh. Basis and window
functions in CT. In 8th International Meeting on Fully Three-Dimensional
Image Reconstruction in Radiology and Nuclear Medicine. Salt Lake City,
Utah, USA, 2005.
[15] P.-E. Danielsson and J. Sunneg˚ardh. Advanced linear modeling and inter-
polation in CT-reconstruction. In 9th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Lin-
dau, Germany, 2007.
[16] B. De Man. Iterative reconstruction for reduction of metal artifacts in com-
puted tomography. Phd thesis, Katholieke Universiteit Leuven, 2001.
[17] B. De Man and S. Basu. Distance-driven projection and backprojection in
three dimensions. Physics in Medicine and Biology, 49(11):2463–2475, 2004.
[18] B. De Man, S. Basu, J.-B. Thibault, J. Hsieh, J.A. Fessler, C. Bouman, and
K. Sauer. A study of four minimization approaches for iterative reconstruc-
tion in X-ray CT. In IEEE Nuclear Science Symposium Conference Record,
volume 5, pages 2708–2710, 2005.
[19] B. De Man, R. Nilsen, and E. Drapkin. High performance image recon-
struction and implementation. In 9th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Lin-
dau, Germany, 2007.
[20] B. De Man, J. Nuyts, P. Dupont, G. Marchal, and P. Suetens. Reduction of
metal streak artifacts in X-ray computed tomography using a transmission
maximum a posteriori algorithm. IEEE Transactions on Nuclear Science,
47(3):977–981, 2000.
[21] B. De Man, J. Nuyts, P. Dupont, G. Marchal, and P. Suetens. An iterative
maximum-likelihood polychromatic algorithm for CT. IEEE Transactions
on Medical Imaging, 20(10):999–1008, 2001.
[22] A. H. Delaney and Y. Bresler. A fast and accurate Fourier algorithm for it-
erative parallel-beam tomography. IEEE Transactions on Image Processing,
5(5):740–753, 1996.
Bibliography 161
[23] I. A. Elbakri and J. A. Fessler. Statistical image reconstruction for polyener-
getic X-ray computed tomography. IEEE Transactions on Medical Imaging,
21(2):89–99, 2002.
[24] H. Erdo˘ gan and J. A. Fessler. Ordered subsets algorithms for transmission
tomography. Physics in Medicine and Biology, 44(11):2835–2851, 1999.
[25] J. A. Fessler and S. D. Booth. Conjugate-gradient preconditioning methods
for shift-variant PET image reconstruction. IEEE Transactions on Image
Processing, 8(5):688–699, 1999.
[26] T. Flohr, K. Stierstorfer, H. Bruder, J. Simon, A. Polacin, and S. Schaller.
Image reconstruction and image quality evaluation for a 16-slice CT scanner.
Medical Physics, 30(5):832–845, 2003.
[27] T. Flohr, K. Stierstorfer, S. Ulzheimer, H. Bruder, A. N. Primak, and C. H.
McCollough. Image reconstruction and image quality evaluation for a 64-
slice CT scanner with z-flying focal spot. Medical Physics, 32(8):2536–2547,
2005.
[28] T. Fuchs. Simulation of noise. As available electronically in May 2009 at
http://www.imp.uni-erlangen.de/phantoms/index.html.
[29] S. Geman and D. E. McClure. Statistical methods for tomographic image
reconstruction. Bulletin of the International Statistical Institute, 52:5–21,
1987.
[30] P. Gilbert. Iterative methods for the three-dimensional reconstruction of
an object from projections. Journal of Theoretical Biology, 36(1):105–117,
1972.
[31] R. Gordon, R. Bender, and G. T. Herman. Algebraic reconstruction tech-
niques (ART) for three-dimensional electron microscopy and x-ray photog-
raphy. Journal of Theoretical Biology, 29(3):471–481, 1970.
[32] P.-C. Hansen. Rank-deficient and discrete ill-posed problems: numerical as-
pects of linear inversion. Society for Industrial and Applied Mathematics,
1997.
[33] K.M. Hanson and G.W Wecksung. Local basis-function approach to com-
puted tomography. Applied Optics, 24(23):4028–4039, 1985.
[34] T. Hebert, R. Leahy, and M. Singh. Fast MLE for SPECT using an inter-
mediate polar representation and a stopping criterion. IEEE Transactions
on Nuclear Science, 35(1):615–619, 1988.
[35] I. A. Hein, K. T. Taguchi, and Silver M. D. Method for helical windmill
artifact reduction with noise restoration for helical multislice CT. US Patent
2006/0029285, February 2006.
[36] G. T. Herman. Image reconstruction from projections. The fundamentals of
computerized tomography. Academic Press, New York, 1980.
162 Bibliography
[37] J. Hsieh. Computed tomography. Principles, design, artifacts and recent
advances. SPIE Press, Washington, 2 edition, 2003.
[38] J. Hsieh, X. Tang, J.-B. Thibault, C. Shaughnessy, and Roy A. Nilsen. Con-
jugate cone beam backprojection for high z-resolution reconstruction. In 8th
International Meeting on Fully Three-Dimensional Image Reconstruction in
Radiology and Nuclear Medicine. Salt Lake City, Utah, USA, 2005.
[39] H.M. Hudson and R.S. Larkin. Accelerated image reconstruction using or-
dered subsets of projection data. IEEE Transactions on Medical Imaging,
13(4):601–609, 1994.
[40] M. Jiang and G. Wang. Convergence studies on iterative algorithms for
image reconstruction. IEEE Transactions on Medical Imaging, 22(5):569–
579, 2003.
[41] P. M. Joseph. An improved algorithm for reprojecting rays through pixel
images. IEEE Transactions on Medical Imaging, 1(3):192–196, 1982.
[42] P. F. Judy. The line spread function and modulation transfer function of a
computed tomographic scanner. Medical Physics, 3(4):233–236, 1976.
[43] M. Kachelriess, M. Knaup, and O. Bockenbach. Hyperfast Perspective Cone-
Beam Backprojection. In IEEE Nuclear Science Symposium Conference
Record, volume 3, pages 1679–1683, October 2006.
[44] M. Kachelriess, S. Schaller, and W. A. Kalender. Advanced single-slice re-
binning in cone-beam spiral CT. Medical Physics, 27(4):754–772, 2000.
[45] A. C. Kak and M. Slaney. Principles of computerized tomographic imaging.
IEEE Press, New York, 1988.
[46] W. A. Kalender. Computed Tomography. Fundamentals, system technology,
image quality, applications. Publicis, Erlangen, 2005.
[47] C. Kamphuis and F. J. Beekman. Accelerated iterative transmission CT re-
construction using an ordered subsets convex algorithm. IEEE Transactions
on Medical Imaging, 17(6):1101–1105, 1998.
[48] A. Katsevich. Theoretically exact filtered backprojection-type inversion al-
gorithm for spiral CT. SIAM Journal on Applied Mathematics, 62(6):2012–
2026, 2002.
[49] T. K¨ ohler. A projection access scheme for iterative reconstruction based on
the golden section. In IEEE Nuclear Science Symposium Conference Record,
volume 6, pages 3961–3965, 2004.
[50] T. K¨ ohler, R. Proksa, and T. Nielsen. SNR-weighted ART applied to
transmission tomography. In IEEE Nuclear Science Symposium Conference
Record, volume 4, pages 2739–2742, 2003.
[51] T. K¨ ohler, H. Turbell, and M. Grass. Efficient forward projection through
discrete data sets using tri-linear interpolation. In IEEE Nuclear Science
Symposium Conference Record, volume 2, pages 15/113–15/115, 2000.
Bibliography 163
[52] J. S. Kole and F. J. Beekman. Evaluation of the ordered subset convex
algorithm for cone-beam CT. Physics in Medicine and Biology, 50(4):613–
623, 2005.
[53] J. S. Kole and F. J. Beekman. Evaluation of accelerated iterative x-ray
CT image reconstruction using floating point graphics hardware. Physics in
Medicine and Biology, 51(4):875–889, 2006.
[54] H. Kunze, K. Stierstorfer, and W. H¨ arer. Pre-processing of projections
for iterative reconstruction. In 8th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Salt
Lake City, Utah, USA, 2005.
[55] L. Landweber. An Iteration Formula for Fredholm Integral Equations of the
First Kind. American Journal of Mathematics, 73(3):615–624, 1951.
[56] K. Lange and R. Carson. EM reconstruction algorithms for emission
and transmission tomography. Journal of Computer Assisted Tomography,
8(2):306–316, 1984.
[57] K. Lange and J. A. Fessler. Globally convergent algorithms for maximum a
posteriori transmission tomography. IEEE Transactions on Image Process-
ing, 4(10):1430–1438, 1995.
[58] G. L. Larson, C. C. Ruth, and C. R. Crawford. Nutating slice CT image
reconstruction apparatus and method. US Patent 5,802,134, 1998.
[59] G. Lauritsch and H. Bruder. Head phantom. As available electronically in
May 2009 at http://www.imp.uni-erlangen.de/phantoms.
[60] R. M. Lewitt. Multidimensional digital image representations using gen-
eralized Kaiser-Bessel window functions. Journal of the Optical Society of
America A, 7(10):1834–1846, 1990.
[61] R. M. Lewitt. Alternatives to voxels for image representation in iterative
reconstruction algorithms. Physics in Medicine and Biology, 37(3):705–716,
1992.
[62] B. Li, G. B. Avinash, and J. Hsieh. Resolution and noise trade-off analysis
for volumetric CT. Medical Physics, 34(10):3732–3738, 2007.
[63] S. Z. Li. MRF Modeling in Computer Vision. Springer Verlag, London,
1995.
[64] M. Magnusson. Linogram and other direct Fourier methods for tomographic
reconstruction. PhD thesis 320, Department of Electrical Engineering,
Link¨ oping University, Link¨ oping, Sweden, November 1993.
[65] M. Magnusson, P.-E. Danielsson, and J. Sunneg˚ardh. Handling of long ob-
jects in iterative improvement of nonexact reconstruction in helical cone-
beam CT. IEEE Transactions on Medical Imaging, 25(7):935–940, 2006.
164 Bibliography
[66] B. P. Medoff, W. R. Brody, M. Nassi, and A. Macovski. Iterative convolu-
tion backprojection algorithms for image reconstruction from limited data.
Journal of the Optical Society of America, 73(11):1493–1500, 1983.
[67] K. Mueller and R. Yagel. Rapid 3-D cone-beam reconstruction with the
simultaneous algebraic reconstruction technique (SART) using 2-D texture
mapping hardware. IEEE Transactions on Medical Imaging, 19(12):1227–
1237, 2000.
[68] K. Mueller, R. Yagel, and J. J. Wheller. Anti-aliased three-dimensional cone-
beam reconstruction of low-contrast objects with algebraic methods. IEEE
Transactions on Medical Imaging, 18(6):519–537, 1999.
[69] K. Mueller, R. Yagel, and J. J. Wheller. Fast implementations of algebraic
methods for three-dimensional reconstruction from cone-beam data. IEEE
Transactions on Medical Imaging, 18(6):538–548, 1999.
[70] F. Natterer. The mathematics of computerized tomography. Wiley and B.G.
Teubner, Stuttgart, 1986.
[71] J. Nocedal and S. J. Wright. Numerical optimization. Springer Series in
Operations Research. Springer, second edition, 1999.
[72] A. Nuttall. Some windows with very good sidelobe behavior. IEEE Trans-
actions on ASSP, 29(1):84–91, 1981.
[73] J. Nuyts, B. De Man, P. Dupont, M. Defrise, P. Suetens, and L. Mortel-
mans. Iterative reconstruction for helical CT: a simulation study. Physics
in Medicine and Biology, 43(4):729–737, 1998.
[74] NVIDIA Corporation. NVIDIA CUDA
TM
programming guide, version 1.1.
As available electronically in May 2009 at http://www.nvidia.com.
[75] NVIDIA Corporation. NVIDIA CUDA
TM
programming guide, version 2.1.
As available electronically in May 2009 at http://www.nvidia.com.
[76] N. Ohyama, S. Ohki, S. Inoue, J. Tsujiuchi, and T. Honda. Discrete Radon
transform in a continuous space. Journal of the Optical Society of America
A, 4:318–324, 1987.
[77] J. D. O’Sullivan. A fast sinc function gridding algorithm for fourier inversion
in computer tomography. IEEE Transactions on Medical Imaging, 4(4):200–
207, 1985.
[78] T. M. Peters. Algorithms for fast back- and re-projection in computed to-
mography. IEEE Transactions on Nuclear Science, 28(4):3641–3647, 1981.
[79] A. Polacin. Design consideration on image reconstruction system for high-
end CT-scanner. In 9th International Meeting on Fully Three-Dimensional
Image Reconstruction in Radiology and Nuclear Medicine. Lindau, Germany,
2007.
Bibliography 165
[80] R. Proksa, R. Proksa, T. Kohler, M. Grass, and J. Timmer. The n-PI-
method for helical cone-beam CT. IEEE Transactions on Medical Imaging,
19(9):848–863, 2000.
[81] J. Radon.
¨
Uber die Bestimmung von Funktionen duch ihre Integralwerte
l¨angs gewisser Mannigf¨altigkeiten. Berichte S¨achsiche Akademie der Wis-
senschaften, 69:262–267, 1917. Reprinted in J. Radon, Gesammelte Abhand-
lungen, Birkh¨ auser Verlag, Vienna, 1987.
[82] G. N. Ramachandran and A. V. Lakshminarayanan. Three-dimensional re-
construction from radiographs and electron micrographs: application of con-
volutions instead of Fourier transforms. Proceedings of the National Academy
of Sciences of the United States of America, 68(9):2236–2240, 1971.
[83] C Riddell, B Bendriem, M H Bourguignon, and J P Kernevez. The ap-
proximate inverse and conjugate gradient: non-symmetrical algorithms for
fast attenuation correction in SPECT. Physics in Medicine and Biology,
40(2):269, 1995.
[84] S. Ryoo, C. Rodrigues, S. Stone, S. Baghsorkhi, S.-Z. Ueng, and W.-M. Hwu.
Program Optimization Study on a 128-Core GPU. Presented at the First
Workshop on General Purpose Processing on Graphics Processing Units,
October 2007.
[85] H. Scharr, S. K¨ orkel, and B. J¨ ahne. Numerische Isotropieoptimierung
von FIR-Filtern mittels Quergl¨ attung. In E. Paulus and F.M. Wahl, edi-
tors, Mustererkennung, volume 97, pages 367–374, Braunschweig, September
1997. Springer.
[86] H. Scherl, B. Keck, M. Kowarschik, and J. Hornegger. Fast GPU-Based CT
Reconstruction using the Common Unified Device Architecture (CUDA). In
Eric C. Frey, editor, IEEE Nuclear Science Symposium Conference Record,
volume 6, pages 4464–4466, 2007.
[87] H. Scherl, M. Koerner, H. Hofmann, W. Eckert, M. Kowarschik, and
J. Hornegger. Implementation of the FDK Algorithm for Cone-Beam CT
on the Cell Broadband Engine Architecture. In J. Hsieh and M. J. Flynn,
editors, Proceedings of SPIE, volume 6510, 2007.
[88] G. Shechter, Th. Kohler, A. Altman, and R. Proksa. The frequency split
method for helical cone-beam reconstruction. Medical Physics, 31(8):2230–
2236, 2004.
[89] L. A. Shepp and B. F. Logan. Reconstructing interior head tissue from x-ray
transmissions. IEEE Transactions on Nuclear Science, 21(1):228–236, 1974.
[90] L. A. Shepp and Y. Vardi. Maximum likelihood reconstruction for emission
tomography. IEEE Transactions on Medical Imaging, 1(2):113–122, 1982.
[91] H. Shi and J.A. Fessler. Quadratic Regularization Design for Iterative Recon-
struction in 3D multi-slice Axial CT. In IEEE Nuclear Science Symposium
Conference Record, volume 5, pages 2834–2836, 2006.
166 Bibliography
[92] R. L. Siddon. Fast calculation of the exact radiological path for a three-
dimensional CT array. Medical Physics, 12(2):252–255, 1985.
[93] E. Y. Sidky and X. Pan. Image reconstruction in circular cone-beam com-
puted tomography by constrained, total-variation minimization. Physics in
Medicine and Biology, 53(17):4777–4807, 2008.
[94] M. Silver, K. Taguchi, I. Hein, B. Chiang, M. Kazama, and I. Mori. Windmill
artifact in multi-slice helical CT. In Proceedings of SPIE - The International
Society for Optical Engineering, volume 5032 III, pages 1918–1927, 2003.
[95] D. L. Snyder and M. I. Miller. The use of sieves to stabilize images produced
with the EM algorithm for emission tomography. IEEE Transactions on
Nuclear Science, 32(5):3864–3872, 1985.
[96] D. L. Snyder, M. I. Miller, Lewis J. Thomas Jr., and D. G. Politte. Noise
and edge artifacts in maximum-likelihood reconstructions for emission to-
mography. IEEE Transactions on Medical Imaging, 6(3):228–238, 1987.
[97] K. Sourbelle. Thorax phantom. As available electronically in May 2009 at
http://www.imp.uni-erlangen.de/phantoms.
[98] S. Srivastava. Accelerated statistical image reconstruction algorithms and
simplified cost functions for x-ray computed tomography. Phd thesis, Univer-
sity of Michigan, Ann Arbour, June 2008.
[99] S. Steckmann, M. Knaup, and M. Kachelriess. Hyperfast general-purpose
cone-beam spiral backprojection with voxel-specific weighting. In IEEE Nu-
clear Science Symposium Conference Record, pages 5426–5433, 2008.
[100] K. Stierstorfer, A. Rauscher, J. Boese, H. Bruder, S. Schaller, and T. Flohr.
Weighted FBP–a simple approximate 3D FBP algorithm for multislice spi-
ral CT with good dose usage for arbitrary pitch. Physics in Medicine and
Biology, 49(11):2209–2218, 2004.
[101] J. Sunneg˚ardh. Iterative enhancement of non-exact reconstruction in cone
beam CT. Master’s thesis, Link¨opings Universitet, September 2004.
[102] J. Sunneg˚ardh. Combining analytical and iterative reconstruction in helical
cone-beam CT. Licentiate Thesis, Link¨opings Universitet, February 2007.
[103] J. Sunneg˚ardh and P.-E. Danielsson. Iterative improvement of non-exact re-
construction in cone-beam CT. In 8th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Salt
Lake City, Utah, USA, 2005.
[104] J. Sunneg˚ardh and P.-E. Danielsson. A new anti-aliased projection operator
for iterative CT reconstruction. In 9th International Meeting on Fully Three-
Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Lin-
dau, Germany, 2007.
Bibliography 167
[105] J. Sunneg˚ardh and P.-E. Danielsson. Regularized iterative weighted filtered
backprojection for helical cone-beam CT. In 9th International Meeting on
Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear
Medicine. Lindau, Germany, 2007.
[106] J. Sunneg˚ardh and P.-E. Danielsson. Regularized iterative weighted filtered
backprojection for helical cone-beam CT. Medical Physics, 35(9):4173–4185,
2008.
[107] J. Sunneg˚ardh and M. Grasruck. Nonlinear regularization of iterative
weighted filtered backprojection for helical cone-beam CT. In IEEE Nu-
clear Science Symposium and Medical Imaging Conference Record, pages
5090–5095, 2008.
[108] K. C. Tam. Three-dimensional computerized tomography scanning method
and system for large objects with smaller area detectors. US Patent
5,390,112, February 1995.
[109] J.-B. Thibault, K.D. Sauer, C.A. Bouman, and J. Hsieh. A three-dimensional
statistical approach to improved image quality for multislice helical CT.
Medical Physics, 34(11):4526–4544, 2007.
[110] H. Turbell. Cone-Beam reconstruction using filtered backprojection. PhD
thesis 672, Department of Electrical Engineering, Link¨ oping University,
Link¨ oping, Sweden, February 2001.
[111] H. K. Tuy. 3D image reconstruction for helical partial cone beam scanners
using wedge beam transform. US Patent 6,104,775, August 2000.
[112] E. Veklerov and J. Llacer. Stopping rule for the MLE algorithm based
on statistical hypothesis testing. IEEE Transactions on Medical Imaging,
6(4):313–319, 1987.
[113] T. E. Walters, W. Simon, D. A. Chesler, and J. A. Correia. Attenuation
correction in gamma emission computed tomography. Journal of Computer
Assisted Tomography, 5(1):89–94, 1981.
[114] G. Wang, Y. Ye, and H. Yu. Approximate and exact cone-beam reconstruc-
tion with standard and non-standard spiral scanning. Physics in Medicine
and Biology, 52(6):R1–R13, 2007.
[115] S. Webb. From the watching of shadows : the origins of radiological tomog-
raphy. Hilger, Bristol, 1990.
[116] F. Xu and K. Mueller. Ultra-fast 3D filtered backprojection on commod-
ity graphics hardware. In IEEE International Symposium on Biomedical
Imaging: Nano to Macro, volume 1, pages 571–574, 2004.
[117] X.-L. Xu, J.-S. Liow, and S. C. Strother. Iterative algebraic reconstruction
algorithms for emission computed tomography: A unified framework and its
application to positron emission tomography. Medical Physics, 20(6):1675–
1684, 1993.
168 Bibliography
[118] G. Z. Yang and D. N. Firmin. The birth of the first CT scanner. IEEE
Engineering in Medicine and Biology Magazine : the Quarterly Magazine of
the Engineering in Medicine & Biology Society, 19(1):120–125, 2000.
[119] A. A. Zamyatin, I. A. Hein, M. D. Silver, and S. Nakanishi. Up-sampling with
Shift Method for Windmill Correction. In IEEE Nuclear Science Symposium
Conference Record, volume 4, pages 2293–2295, 2006.
[120] A. A. Zamyatin, A. Katsevich, M. D. Silver, and S. Nakanishi. Helical CT
reconstruction with large cone angle. In IEEE Nuclear Science Symposium
Conference Record, volume 4, pages 2264–2267, 2006.
[121] W. Zbijewski. Model-based image reconstruction in X-ray Computed Tomog-
raphy. Phd thesis, Utrecht University, Utrecht, Juni 2006.
[122] G.L. Zeng and G.T. Gullberg. Unmatched projector/backprojector pairs
in an iterative reconstruction algorithm. IEEE Transactions on Medical
Imaging, 19(5):548–555, 2000.
[123] K. Zeng, Z. Chen, L. Zhang, and G. Wang. An error-reduction-based algo-
rithm for cone-beam computed tomography. Medical Physics, 31(12):3206–
3212, 2004.
[124] A. Ziegler, T. K¨ ohler, T. Nielsen, and R. Proksa. Iterative cone-beam CT
image reconstruction with spherically symmetric basis functions. In 8th
International Meeting on Fully Three-Dimensional Image Reconstruction in
Radiology and Nuclear Medicine. Salt Lake City, Utah, USA, 2005.
[125] A. Ziegler, T. K¨ ohler, T. Nielsen, and R. Proksa. Efficient projection and
backprojection scheme for spherically symmetric basis functions in divergent
beam geometry. Medical Physics, 33(12):4653–4663, 2006.
[126] A. Ziegler, T. K¨ ohler, and R. Proksa. Noise and resolution in images recon-
structed with FBP and OSC algorithms for CT. Medical Physics, 34(2):585–
598, 2007.
[127] A. Ziegler, T. Nielsen, and M. Grass. Iterative reconstruction of a region
of interest for transmission tomography. Medical Physics, 35(4):1317–1327,
2008.
[128] Y. Zou and X. Pan. Exact image reconstruction on PI-lines from minimum
data in helical cone-beam CT. Physics in Medicine and Biology, 49(6):941–
959, 2004.
[129] Y Zou, E. Y. Sidky, and X. Pan. Partial volume and aliasing artefacts in
helical cone-beam CT. Physics in Medicine and Biology, 49(11):2365–2375,
2004.

Link¨ping Studies in Science and Technology o Dissertation No. 1264 Author Johan Sunneg˚ ardh Department of Electrical Engineering Link¨ping University o SE-581 83 Link¨ping, Sweden o Copyright c 2009 Johan Sunneg˚ ardh Sunneg˚ ardh, Johan Iterative Filtered Backprojection Methods for Helical Cone-Beam CT ISBN 978-91-7393-586-9 ISSN 0345-7524
A Typeset with L TEX Printed in Sweden by LiU-tryck, Link¨ping, 2009 o

iii

Abstract
State-of-the-art reconstruction algorithms for medical helical cone-beam Computed Tomography (CT) are of type non-exact Filtered Backprojection (FBP). They are attractive because of their simplicity and low computational cost, but they produce sub-optimal images with respect to artifacts, resolution, and noise. This thesis deals with possibilities to improve the image quality by means of iterative techniques. The first algorithm, Regularized Iterative Weighted Filtered Backprojection (RIWFBP), is an iterative algorithm employing the non-exact Weighted Filtered Backprojection (WFBP) algorithm [Stierstorfer et al., Phys. Med. Biol. 49, 22092218, 2004] in the update step. We have measured and compared artifact reduction as well as resolution and noise properties for RIWFBP and WFBP. The results show that artifacts originating in the non-exactness of the WFBP algorithm are suppressed within five iterations without notable degradation in terms of resolution versus noise. Our experiments also indicate that the number of required iterations can be reduced by employing a technique known as ordered subsets. A small modification of RIWFBP leads to a new algorithm, the Weighted Least Squares Iterative Filtered Backprojection (WLS-IFBP). This algorithm has a slightly lower rate of convergence than RIWFBP, but in return it has the attractive property of converging to a solution of a certain least squares minimization problem. Hereby, theory and algorithms from optimization theory become applicable. Besides linear regularization, we have examined edge-preserving non-linear regularization. In this case, resolution becomes contrast dependent, a fact that can be utilized for improving high contrast resolution without degrading the signal-tonoise ratio in low contrast regions. Resolution measurements at different contrast levels and anthropomorphic phantom studies confirm this property. Furthermore, an even more pronounced suppression of artifacts is observed. Iterative reconstruction opens for more realistic modeling of the input data acquisition process than what is possible with FBP. We have examined the possibility to improve the forward projection model by (i) multiple ray models, and (ii) calculating strip integrals instead of line integrals. In both cases, for linear regularization, the experiments indicate a trade off: the resolution is improved at the price of increased noise levels. With non-linear regularization on the other hand, the degraded signal-to-noise ratio in low contrast regions can be avoided. Huge input data sizes make experiments on real medical CT data very demanding. To alleviate this problem, we have implemented the most time consuming parts of the algorithms on a Graphics Processing Unit (GPU). These implementations are described in some detail, and some specific problems regarding parallelism and memory access are discussed.

iv .

som i RIWFBP. En nackdel ar dock att brusniv˚ ¨kar j¨mf¨rt a a ¨ an o a o med RIWFBP. I avhandlingen f¨resl˚ en tredje metod. vilket betyder att de rekon¨ a o stuerade bilderna i varje iteration filtreras med ett icke-linj¨rt kantbevarande filter. Med. Resultaten visar en tydlig minskning av artefaktniv˚ an under de f¨rsta fyra iterationerna. bygger p˚ iterativ o a o a till¨mpning av den icke-exakta metoden WFBP [Stierstorfer et al. som o as o med en gradientbaserad metod minimerar en kvadratisk m˚ alfunktion. En nackdel med RIWFBP ar att den till skillnad fr˚ de flesta andra iterativa ¨ an rekonstruktionsmetoder inte enkelt kan beskrivas som l¨sningen till ett minimero ingsproblem. Phys. o o Till skillnad fr˚ en vanlig r¨ntgenunders¨kning som anv¨nder r¨ntgenstr˚ i en an o o a o alar enda riktning. som p˚ engelska f¨rkortas RIWFBP. a men har f¨rdelen att den kan analyseras och utvidgas med v¨lk¨nda verktyg fr˚ o a a an omr˚ adet optimeringsl¨ra. F¨r att f¨rb¨ttra f¨rh˚ a o o a o allandet mellan sk¨rpa och a brus har aven icke-linj¨r regularisering unders¨kts. f¨rkortad WLS-IFBP. kan en delm¨ngd anv¨ndas.v Popul¨rvetenskaplig sammanfattning a Datortomografi (f¨rkortat DT eller CT). 2004]. a Med avseende p˚ sk¨rpa kontra brus ar den linj¨rt regulariserade RIWFBPa a ¨ a metoden likv¨rdig med WFBP. o o a n˚ agot n¨r denna teknik anv¨nds. a an a Alla rekonstruktionsmetoder som idag anv¨nds f¨r medicinsk datortomografi a o ar icke-exakta. anv¨nder CT r¨ntgenstr˚ i flera olika riktningar f¨r att rekona o alar o struera bilder av skikt i kroppen. a Biol. ders¨ks olika iterativa metoder f¨r att undertrycka dessa artefakter. anv¨nda alla tillg¨ngliga m¨tningar i varje a o a a a a a iteration. En s˚ kallad rekonstruktionsmetod erfodras f¨r a o att kunna ber¨kna bilderna fr˚ uppm¨tta data. a a Den f¨rsta metoden.. 2209-2218. Ist¨llet f¨r att. o as o a a vilket inneb¨r att de rekonstruerade bilderna i varje iteration filtreras med ett a linj¨rt l˚ a agpassfilter. bruso niv˚ samt spatiell uppl¨sning m¨tts genom rekonstruktion av f¨r ¨ndam˚ aer o a o a alet konstruerade testobjekt. 49. Detta utg¨r grunden till metod nummer o tv˚ f¨rkortad OS-IWFBP. Det visar sig att antalet n¨dv¨ndiga iterationer sjunker a. F¨r RIWFBP har reduktion av artefakter. samt f¨rb¨ttra o o o a de rekonstruerade bilderna med avseende p˚ sk¨rpa och brus. men p˚ nyare mer effektiva tomografer med ¨ o a 64 eller fler detektorrader kan st¨rande artefakter uppst˚ I denna avhandling uno a. a Experiment visar att denna typ av regularisering kan f¨rb¨ttra den sk¨rpan f¨r o a a o h¨gkontrastdetaljer (ben) utan f¨rs¨mring av signal/brus-f¨rh˚ o o a o allandet i omr˚ aden . En negativ bieffekt ar att bildernas brusniv˚ o ¨ a okontrollerat okar med antalet iterationer. ar en vanligt o ¨ o ¨ f¨rekommande teknik inom medicinsk bilddiagnostik f¨r att avbilda kroppens inre. F¨r ¨ldre tomografer med f˚ o o a a detektorrader ar dessa fel f¨rsumbara. Den nya metoden undertrycker artefakter ungef¨r lika effektivt som RIWFBP-metoden. aven kallad skiktr¨ntgen. F¨r att f¨rb¨ttra konvergensegenskaper ¨ o o a och undvika denna bieffekt f¨resl˚ och unders¨ks s˚ kallad linj¨r regularisering. Detta betyder att de framr¨knade bilderna inneh˚ ¨ a aller fel oavsett noggrannhet och uppl¨sning vid datainsamlingen.

P˚ grund av stora datam¨ngder ar experiment p˚ riktiga medicinska proa a ¨ a jektionsdata mycket kr¨vande. Med hj¨lp av icke-linj¨r regularisering kan denna a a kompromiss delvis undvikas.vi med l˚ kontraster (mjukdelar). . Via dare diskuteras specifika problem r¨rande parallelisering och minnes˚ o atkomst. o a samt systemets r¨relse under datainsamlingen. aga Till skillnad fr˚ direkta metoder m¨jligg¨r iterativa metoder s˚ an o o asom RIWFBP och WLS-IFBP modellering av r¨ntgenk¨llan och detektorelementens storlekar. I det linj¨rt a regulariserade fallet visar resultaten att noggrann modellering ger b¨ttre sk¨rpa a a till priset av okade brusniv˚ ¨ aer. I avhandlingen beskrivs dessa implementationer samt ungef¨rliga prestandavinster. F¨r att minska ber¨kningstiderna har de mest a o a ber¨kningskr¨vande operationerna implementerats p˚ grafikh˚ a a a ardvara. Vi har j¨mf¨rt rekonstruerade o a o bilder med och utan noggrann modellering av dessa egenskaper.

Ulrike for all the love and support you have given me. for proof-reading. for inviting me to work and study at Siemens in Forchheim. and for always leaving his door open. Maria Magnusson for interesting discussions. All other people that I learned to know at Siemens during my stay in Germany. Thomas Flohr. for all advice concerning the writing. Maria Antonson and Petter Quick. and for many interesting technical discussions. Fredrik Larsson. for providing simulation and reconstruction software. and for being an invaluable source of information about CT. The financial support from Siemens Healthcare is gratefully acknowledged. Siemens Healthcare. and for many interesting discussions. for proof-reading. for many valuable comments and questions regarding my work. . Siemens Healthcare. All past and present members at the Computer Vision Laboratory for creating a friendly and inspiring atmosphere.vii Acknowledgments I would like to take this opportunity to thank all those who have contributed to this thesis. My family for all support during the writing of this thesis. for making me feel welcome. and Klas Nordberg for proof-reading and providing valuable questions and comments. and for providing answers to my never ending flow of questions regarding various topics in the field of image processing and computed tomography. Siemens Healthcare. for providing excellent photos of a modern CT system. for many interesting discussions. Michel Defrise for many useful questions and comments on my licentiate thesis. for his never ending commitment and enthusiasm. Tommy Elfving and Lars Eld´n for sharing their knowledge in numerical analysis e and optimization theory. Karl Stierstorfer. Michael Felsberg for all his support and advice in the final stages of my work. G¨sta Granlund for welcoming me to work and study at the Computer Vision o Laboratory. First and foremost. Germany. Per-Erik Danielsson who introduced me to the field. and for making me happy. CMIV. Herbert Bruder. and for proof-reading the manuscript. Johan Hedborg.

viii .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . 3. . . . . . .4 Experiments . . . . . . . . . . 5. . . . . 1 1 2 4 5 7 8 9 9 13 19 19 21 23 23 25 26 30 32 35 35 37 39 45 45 48 49 50 50 57 60 63 65 2 Analytical reconstruction methods 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . .4. 2. . .4 Local region of interest . . . . . . . .5 Contributions and publications . . . . . . . . . . .4. . . .1 Artifacts and noise measurements . 5. . . . .Contents 1 Introduction 1. . . . . . . . . . . . . . . . . 5. . . . . . . . .3.3 Irradiation functions . . 5. . . . . . . . . . . .2 Weighted filtered backprojection . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . 3 Iterative reconstruction methods 3. 5. Long objects. . . . . 4 Simulation and evaluation 4. . . .3 Local region of interest (LROI) . . . . . 4. .1 Two-dimensional image reconstruction . . .1 Description of the reconstruction algorithm .2 Reconstruction methods and artifacts . .2 Spatial resolution . . . Application overview. . . . . . .1 Matrix representation . . . . . . .3 Higher number of iterations. . . . . . . .1 A basic example. . 1. . . . . . . . . . . . . . . . .5 Discussions and conclusions . . . . . . . . . . 3. . . . . . . . . . . . . . . . 4. . . . 3. . . . . . . . . . . . . . . . . . . . . . . . 3. . . .4 The Joseph forward projection method . . . . . . . . .3 Iterative Filtered Backprojection (IFBP) 1. . . . . . . 5. . . . .3 The projection operator . . . . . . . . . .3. . . . . . .1 Computed Tomography (CT) . . . . . . .2 Phantoms and error/noise measures . . . . . . . . . .3 Spatial resolution measurements . . . . . . . . . . . . .4 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4.2 Different types of iterative methods . . . . . . . . . . . . 5. . . . . . . . .6 Notation . . . . . . . . . . . . . . . . . . .4 Regularization . . . . . . . . 1. . . . . . . . . . . . . . . . . . 3. . . . . . . . . 3. . . . . . . . . . . . . . . .4. . . . . . . . . . . . . .2 The regularization operator R . . . . . . . . . . . . . 1. .1 Generation of input projection data . . . . . . . . . . . . . .2 Basis functions . . . . . . . . . . 5 Regularized iterative weighted filtered backprojection 5. 1. . . . . . . . . . . . . . . . . . . . . . . . .

. . . .3. . . . . . . 9. . . . . . . . . . . . . .2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . .2 The Ordered Subsets (OS) technique . . . . . . . . . .3. . . . . . 10. . . . . . . . . .1 Method description . . 8. . . . . . . . . . . . 8. . . . . . . . . .2 Edge versus surface based resolution measurements 8. . . . . . 10. . . . . 9 Data acquisition modeling 9. . . . . . .x 6 Ordered Subsets IWFBP (OS-IWFBP) 6. . 7. .2 Weighted least squares IFBP (WLS-IFBP) . . . . . . . . . . . 10. . . . . . 8. . . . . . . . . . . . . .3 OS-IWFBP . . . 9. . . . . . . . . .2. . . . . .3. . . . . . .3 Artifacts and noise . . . . . . . . 10. . . . 8. . . . . . . . . . . . . .2 Spatial resolution and noise . . . . . . . . . . . . . . . . . .1 Irradiation functions . . . . . . . . . . . . . . .4 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . 10. . . . . . . . . . . . . .3. . 10 Implementation on Graphics Processing Units 10. . . . . . . . . . . . . . . . . . . . . . . .6 Discussions and conclusions . . . . . . . . . . . . . . .1 An improved acquisition model 9. . . . .3. . . . . . . . . 7. . . . . .3 A strip integral projection method . . . . . . . . . . . . . . . . .1 Kernels and threads . . . 10. . 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. . . . . . . . . . . . . 7. . . . . . . . .2 Experiments . . . . . . . . . . . .3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. . . . . . 10. . 95 . . . . . .4 Constant memory . . . . . . . . . .3. . . .2 The GPU from a CUDA perspective . . . . . . . . 108 113 113 114 114 117 123 127 127 130 131 135 135 136 138 138 139 142 142 143 145 145 147 149 . . . . . . 7. . 7 Relation between RIWFBP and least squares methods 7. . . . . . . 10. . . .1 Artifact reduction . . . . . . . 9. . . . . . . . . . . . .3. . . 98 . .1 Introduction .1 Spatial resolution versus local contrast . . . . . . . . . . . . 9.3 Experiments . . . . . . . . 6. . . . . . . . . . . . 8. . . . . 6. . . . . .2. . . . . . . 104 . . 98 . . . . . . . . . . . . . . . . . . . .3 Discussions and conclusions . . . . . . . . . (GPU) .3 Discussions and conclusions .1 The WFBP/standard backprojector . . . . . . . . . . . . . . .5 Backprojection . . . . . . . . . 10. . 9. 10. . . . . . . . .3. . . . . . . . . . . .1 Motivation . . . . . . . . .2. . . . . . . . . .2 The Joseph backprojector PT . . . . 7. . . . . . . . . . . . . . . . . . . . . .5. . . . . . . . . . . .4 Forward projection . . . . . .3 Programming CUDA . . .1 Introduction . . 8 Nonlinear regularization 8. .3. . .4 Discussions and conclusions . . . . . . . . . . . . . . . . . . . . . . . .2 Multiple ray forward projection . . . . . . . . . . . . . . .4 Experiments . . . . . . . . . .2 Description of the reconstruction method . . 95 . . . . .5. . . . . . . . . . . . . . . . . . .2 Thread organization . .1 Introduction . . . . . 95 . . . . . . . . .3 Textures . . . .5 Discussions and conclusions . . . . . . . . . . . 6. . . . . . .3. . . . . . . . . . . . . . . . . . . . 6. . 101 . . . . . . . . . . . . . . . . . . . . Contents 71 71 71 73 73 79 83 83 84 87 87 88 90 . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . .1 Summary of conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 References 158 . . . . . . . . . . . . . . 151 11. . . . . . 155 B Notation .2 Suggestions for future research . . . . . . . . . . .Contents xi 11 Conclusions and future research 151 11. 153 Appendices 155 A Acronyms . . . . . . . . . . . . . . . . . .

xii Contents .

several major improvements have been made in this direction. it is of great interest to be able to reproduce images of the interior of objects.1 Computed Tomography (CT) In the areas of medical diagnostics and non-destructive testing. Here. The technology became a huge success and in 1979. Fig. no cables need to be attached to the gantry. A CT scanner uses digitally sampled X-ray images acquired in multiple directions to calculate cross-sectional images of the X-ray attenuation of an object. the slip ring connected gantry is rotating with a constant speed of between one and four turns per second. Alan Cormack and Hounsfield were awarded the Nobel Prize for this invention. Thanks to the so called slip ring technique for transfer of electrical power and data. when Sir Godfrey Hounsfield at Electrical and Musical Industries limited (EMI) patented the first CT scanner [118]. different scanning trajectories are obtained. The data acquisition system. . Later developments are covered in the books by Kalender [46] and Hsieh [37]. One common technique to accomplish this feat is known as Computed Tomography (CT). obtained by translating the patient table with constant speed during data acquisition. we give a brief presentation of the state of the art at the time of writing. The most common one is the helical trajectory.e. During an examination. the gantry was supported by cables which typically tolerated one turn after which the gantry had to change its direction of rotation.Chapter 1 Introduction 1. Previously. Since 1972. 1.1 shows a photograph of a modern CT scanner. orthogonal to the gantry rotation plane. Depending on this motion. i. is attached to the gantry hidden inside the main part of the scanner. A comprehensive survey of early developments in CT is given by Webb [115]. Different regions of the patient are irradiated by translating the patient table along its z-axis. High acquisition speed is generally desirable since this reduces examination time and image artifacts due to motion. which originates from 1972. consisting of an X-ray source and a detector.

An analytical method is based on exact or approximate mathematical inversion of the projector Pphys that models the data acquisition process. Reconstruction methods for CT can be either analytical or iterative. (1. the cone-angle κmax . and three years later. In contrast. it is now possible to perform a full body examination in less than five seconds. In combination with other techniques for accelerating data acquisition. many developments towards higher acquisition speeds have been introduced by increasing the number of detector rows and improving systems for transferring data from the gantry. or equivalently. The detector is shaped as a cylinder centered around the source and aligned with the z-axis.2 Reconstruction methods and artifacts An important part of any CT system is the reconstruction method. During the past decade. 1. In 1998. 64 row detectors are quite common. but use an iterative procedure to generate a sequence of reconstruction results converging to . the number of detector rows in high end scanners was increased to sixteen [46]. which transforms the measured data into images. iterative methods do not involve explicit inversion. Its height h when projected onto the center of rotation. Today. the first scanners with four detector rows were released.4h = 1.2 illustrates the helical scanning geometry in the coordinate system of the examined object.1: A modern medical CT scanner (by courtesy of Center for Medical Image Science and Visualization (CMIV). Sweden).1) where RS is the distance from the source to the center of rotation. such as dual source systems. o Fig. 1. Link¨ping. determines an upper limit on the helical pitch P approximately equal to Pmax ≈ 1. in 2001.2 Introduction z Figure 1.4 × 2RS tan κmax .

1.2 Reconstruction methods and artifacts

3

Projection of the detector onto the center of rotation (isocenter)

z

Detector

h

P

2κ ma

x

y X-ray source

x Figure 1.2: Scanning geometry for helical cone-beam CT. To simplify the visualization, the detector height h and the pitch P of the helix are exaggerated. an image or volume f that in some sense minimizes the difference between the projection data Pphys f and the measured data. These things will be explained more precisely in Chapter 2 and 3. Although the EMI scanner by Hounsfield employed an iterative method [118], reconstruction methods in commercially available CT systems are mainly analytical. One reason for this is that analytical methods are computationally much less expensive than iterative methods. Analytical methods can be either exact and non-exact. Exact methods are based on exact inversion of the projector Pphys in the continuous domain. Thus, in an ideal scanning situation (monochromatic radiation, linear detector response, and no geometric inaccuracies), reconstruction results can be made arbitrarily accurate by using sufficiently high sampling densities and tube currents. Although efficient exact methods based on work by Katsevich [48] and Zou and Pan [128] exist, they are currently not found in clinical use. Instead, manufacturers of clinical CT systems employ non-exact methods, or more specifically, Filtered Backprojection (FBP) methods [58, 44, 111, 26, 100]. Due to approximations in the derivation of these methods, reconstruction results are contaminated by cone artifacts, the name stemming from the fact that these artifacts become more pronounced for higher cone angles. In return, non-exact methods are computationally less demanding, simpler to implement, and offer a better dose utilization than exact methods. One state-of-the-art analytical non-exact reconstruction method in helical

4

Introduction

y x Figure 1.3: An axial image of a clock phantom reconstruction distorted by cone and windmill artifacts. The cone artifacts are slowly varying, and appear as dark and bright regions in the vicinity of the spheres. The windmill artifacts appear as multiple dark and bright streaks tangent to the spheres. Greyscale window ±20 Hounsfield Units (HU). cone-beam CT is the Weighted Filtered Backprojection (WFBP) by Stierstorfer et al. [100]. Fig. 1.3 shows an axial image of a WFBP reconstruction of the sphere clock phantom by Turbell [110], specially designed for examining the nonexactness of a reconstruction method. The image shows two types of artifacts: windmill artifacts and cone artifacts. Wind-mill artifacts are caused by aliasing in the z-direction of the detector, and look like alternating dark and bright streaks extending radially from the upper and lower ends of the spheres. The cone-artifacts are more slowly varying than the windmill artifacts. In Fig. 1.3, they show up as one or two pairs of dark and bright regions extending from the spheres. Another example of cone artifacts and windmill artifacts is given in Fig. 1.4, showing a sagittal image of a WFBP reconstruction of a physical head phantom.

1.3

Iterative Filtered Backprojection (IFBP)

The purpose of the work presented in this thesis is to investigate the possibilities to improve FBP methods with respect to artifact reduction, spatial resolution, and signal to noise ratio, by means of Iterative Filtered Backprojection (IFBP). Below, we give a brief presentation of a simple IFBP method. A more detailed presentation and analysis is given in Chapter 5 and 7.

1.4 Thesis outline

5

z y Figure 1.4: A sagittal image of a physical head phantom reconstruction distorted by cone and windmill artifacts. The dashed circle and ellipse indicate windmill artifacts, and the solid ellipse indicates a cone artifact. Fig. 1.5 illustrates the basic principle of IFBP. Suppose that the input data represented by the vector pin ∈ RM have been obtained from the object of interest f ∈ L2 (R3 ) by the linear operator Pphys : L2 (R3 ) → RM modeling the data acquisition process. Furthermore, let the matrix Q ∈ RN ×M represent a non-exact FBP method (N is the number of reconstructed voxels). An approximation of the correct result is then given by f0 = Qpin ∈ RN . By employing a discrete model P ∈ RM ×N of the physical projection operator Pphys , projection data p0 = Pf0 of the approximative result can be calculated. Now, the difference pin − p0 can be used to create an approximate correction image fdiff = Q(pin − p0 ). Finally, addition of f0 and fdiff yields a corrected result f1 , which under certain circumstances contains less artifacts than f0 . By repeating this procedure, a sequence of gradually improved results f0 , f1 , ... is obtained. In the field of CT, IFBP methods have successfully been applied to problems where data are missing due to heavy objects [66] and due to an incomplete source trajectory [123]. Danielsson and Magnusson [12] proposed to use this scheme to suppress artifacts caused by non-exact reconstruction in helical cone-beam CT with high cone angles.

1.4

Thesis outline

Chapter 2: Analytic reconstruction methods. Some basic concepts in CTreconstruction such as ramp filtering and backprojection are presented, followed by a short description of the WFBP method. Chapter 3: Iterative reconstruction methods. As an introduction we describe the well-known Landweber method [55], which is one of the simplest methods for iterative reconstruction, followed by examples and comments on various parameters and alternative methods. In Section 3.3, we discuss the concepts of basis and irradiation functions. This is followed by an introduction of regulariza-

5: Iterative Filtered Backprojection (IFBP) tion techniques for least square methods. and the possibility to iteratively reconstruct a region of interest that is partially or completely contained by the examined object. Here. Chapter 7: Relation between RIWFBP and least squares methods. The methods for assessing spatial resolution and measuring noise and errors are also presented. we describe how simulated projection data will be generated in the subsequent chapters.8◦ and κmax = 9. Ordered Subsets (OS) is a well-known technique for reducing the number of required iterations. leading to the . Besides a description and analysis of the iterative method. we present the OS technique in general and show how it can be applied to the RIWFBP method. between least squares methods and IFBP methods suggests a way to introduce regularization with penalty terms for the IFBP methods. Subsequent sections comprise experimental studies of artifact suppression. which is one of the main contributions in this thesis. In this chapter. Geometrical parameters and the different phantoms are explained and specified. Experiments with high cone angles (κmax = 4.6◦ ) show how OS-IWFBP outperform RIWFBP with respect to cone artifact suppression in these cases. In this chapter. and local region of interest reconstruction. and sometimes even equivalence. Because of an advanced normalization procedure in the WFBP method. Chapter 6: Ordered Subsets IWFBP (OS-IWFBP).6 Introduction     Figure 1. the RIWFBP method is not easily expressed in terms of optimization theory. spatial resolution versus noise. behavior over multiple iterations. This chapter starts with a description of the Regularized Iterative Weighted Filtered Backprojection (RIWFBP) method. a simplification of the normalization in the RIWFBP is presented. we discuss the design of the regularization operator. Chapter 4: Simulation and evaluation. Chapter 5: Regularized Iterative Filtered Backprojection (RIWFBP). The motivation behind this is that the similarity.

Evaluation by visual inspection. Noise and spatial resolution measurements at different contrast levels show that nonlinearly regularized WLS-IFBP can be used to improve high contrast spatial resolution and reduce artifacts while increasing the signal to noise ratio for low-contrast regions. we investigate the effects of accurate multiple ray models of the geometry during data acquisition.1.e. 1. but with a slightly slower rate of convergence than for RIWFBP. The forward projection and backprojection operations in CT reconstruction are computationally very expensive.5 Contributions and publications Much of the material in this thesis has been adapted from previous publications. Chapter 8: Nonlinear regularization. A careful choice of the forward projection operator P is important for obtaining good results in iterative reconstruction methods. and root mean squared error measurements show that WLS-IFBP produces images very similar to those of RIWFBP.5 Contributions and publications 7 Weighted Least Squares IFBP (WLS-IFBP) method. In return. ch. we describe how nonlinear regularization. This chapter provides a summary of conclusions and presents some questions remaining to be answered. detector. An earlier version without regularization or evaluation of spatial resolution versus noise was first published in [103]. The second part of this chapter presents a new strip integrating projection operator that can reduce aliasing related artifacts and improve spatial resolution. as well as experiments with data from a . Chapter 9: Data acquisition modeling. Further improvements of the method. In this chapter. In [102. It is shown that accurate modeling of source. and reduce artifacts related to aliasing. This is followed by experiments with WLS-IFBP and the qGeneralized Gaussian Markov Random Field (q-GGMRF) potential function by Thibault et al. Chapter 10: Implementation on Graphics Processing Units (GPU). regularization and spatial resolution versus noise were examined. Chapter 11: Conclusions and future research. First. the WLS-IFBP method is easier to analyze and improve using existing theory from the field of optimization theory. [109]. i. and how they are related to existing publications. can be used with the RIWFBP and WLS-IFBP.. 5] and [105]. In the first part of this chapter. The presentation and evaluation of RIWFBP method in Chapter 5 is one of the major contributions in the thesis. we describe our experiences in accelerating these operations by implementing them on graphics hardware using the NVIDIA Compute Unified Device Architecture (CUDA). noise measurements. Below follows a brief summary of the contributions. and gantry rotation can improve spatial resolution. non-quadratic potential functions.

W . • Scalars and scalar-valued functions are denoted with italic letters. The superscript T is used for denoting the transpose of a vector or a matrix. corrections and improvements of the theory and reconstruction methods have been made.g. The jth component of a vector is written as rj . an element on the ith row and jth column of a matrix P is denoted by pij . x. Below. • General linear operators are denoted by italic uppercase letters. A subscript on these symbols means weighting according to x 2 = x. and has not yet been published. we explain the basic mathematical conventions used in the thesis. • Vectors and matrices are denoted by bold lowercase and uppercase letters respectively. modeling can affect reconstruction results. e. the evaluation have been improved by including assessment of contrast dependent spatial resolution and reconstruction of data from a clinical scanner. e. and f (r). The first part of Chapter 9 examines effects of more accurate forward projection models and effects of backprojector type. Furthermore.g. Since then.g. r and P. x W = xT Wx.8 Introduction clinical scanner were published in [106]. Correspondingly. but computationally expensive. The WLS-IFBP method in Chapter 7 and nonlinear regularization in Chapter 8 were initially published in [107]. f (x). The second part of Chapter 9 presents a strip integration technique for aliasing suppressed forward projection and backprojection [104]. Pphys . and what one can expect from good approximations of the models examined here.6 Notation A selection of symbols and a list of acronyms are found in the appendices. · and · respectively. • Euclidean scalar products and norms are denoted by ·. This contribution provides insight in how accurate. 1. e.

The presentation is mainly inspired by the more extensive text by Kak and Slaney [45]. y).1 that illustrates an x-ray beam with initial photon intensity I0 traveling through an object described by the linear attenuation function f (x. θ) = − log I(t. θ) = I0 e − ∫ Γt.Chapter 2 Analytical reconstruction methods 2. Any reconstruction method for two-dimensional parallel beam CT faces the problem of inverting this transform. Consider Fig. (2.θ f (x. It provides a connection between the Radon transform and the Fourier transform of a two-dimensional function: . θ) is given by I(t. After traveling through the object.θ f (x. θ) is known as the Radon transform. and works by Magnusson [64] and Turbell [110].1 Two-dimensional image reconstruction The mathematical problem of two-dimensional image reconstruction from line integrals was first solved by Johann Radon in 1917 [81].2) The linear operator R mapping f (x. 2. y)dl. (2. the Fourier slice theorem is very useful. Kak and Slaney [45] and Natterer [70]. θ) I0 = Γt. Here we present the reconstruction problem in general followed by a closer look at the widely used filtered backprojection (FBP) method by Bracewell and Riddle [6]. and Ramachandran and Lakshminarayanan [82]. In order to invert the Radon transform.y)dl . see for instance the books by Herman [36]. Several text books on the subject exist.1) Taking the logarithm yields p(t. the remaining photon intensity I(t. Practical solutions suitable for computer implementations have been an active research topic since early 1970s. y) to p(t.

2 (2. θ) = (Ft p)(ρ.10 Analytical reconstruction methods      Figure 2.y)dl . Theorem 1 (The Fourier slice theorem) The two-dimensional Fourier transform Ff of a function f (x.3) The theorem is illustrated in Fig.1: Photon intensities of attenuated and non-attenuated x-ray beams ∫ differ by a factor e − Γt. Performing a final inverse two-dimensional Fourier transform yields the reconstructed result. Theorem 1 can be applied directly to parallel projection data by computing the two-dimensional Fourier transform of f (x. ρ sin φ) = (Ft (Rf ))(ρ. φ = θ − π . θ). However.             Figure 2. 2. y) through a series of one-dimensional Fourier transforms Ft (Rf ). applying Theorem 1 directly is associated with the problem of mapping radially sampled Fourier data to a .2: Illustration of the Fourier slice theorem. y) along a radial line equals the one-dimensional Fourier transform along the radial direction of the Radon transform: (F2 f )(ρ cos φ.2 and a formal proof is given by Kak and Slaney in [45].θ f (x.

Projection and backprojection are illustrated in Fig. θ)dθ (2. y) decays as 1/d. Then. Fig. these Direct Fourier Methods (DFMs) have been less popular than filtered backprojection (FBP). An FBP method implements the inversion formula π f (x.3.1 Two-dimensional image reconstruction 11 Cartesian grid.φ+π/2) = 0 (Ft p)(ρ. v)ei2π(xu+yv) dudv = = (u = ρ cos φ. This step is followed by a backprojection. 2. θ)dθ . θ)|ρ|ei2πρ(x sin θ−y cos θ) dρdθ = 0 π −1 (Ft (Ft p · Fg∞ ))(x sin θ − y cos θ. this image will be non-zero not only on the support of the original object. y) = −∞ (Ff )(u. backprojection is the Hilbert adjoint operator R∗ to R [70]. θ) ∗ g∞ )(x sin θ − y cos θ. the resulting image fbp (x. θ)dθ = 0 (p(·. θ). the Since the 2D Fourier transform of this function is 1/ u2 + v 2 . 2. also gives an intuitive explanation for the Point Spread Function (PSF) of the combined operation projection followed by backprojection. the projection data p = Rf are filtered with a linear and spatially invariant so called rampfilter g∞ in the radial (t) direction. but also in the rest of the image. According to Theorem 1. ∞ ∞ −∞ ∞ −∞ ∞ −∞ π f (x.4) can be derived mathematically as follows.3. inspired by Magnusson [64]. Even if solutions to this problem exist (see Magnusson [64] and O’Sullivan [77]). In fact. dudv = |ρ|dρdφ) π/2 −π/2 π (Ff )(ρ cos φ.4) As a first step. y) = 0 (p(·. y) is a good approximation of the point spread function that we search for. filtering backprojection (2. we may conclude that fbp (x. Equation (2. v = ρ sin φ. θ) ∗ g∞ )(x sin θ − y cos θ. the inverse filter √ should look like u2 + v 2 in the Fourier domain.5) . this filtering can be implemented as a one-dimensional filtering in the radial direction of the projection data p(t.2. where d is√ distance to the object. In terms of operator theory. Clearly. Assume that projections of a very small object have been collected and are being backprojected as shown to the right in Fig. 2. which consists of adding (in the case of a finite number of angles) the filtered projection data to a digital image along the rays from which the projection data has its origin.3. ρ sin φ) ei2πρ(x cos φ+y sin φ) |ρ|dρdφ =(Ft p)(ρ. by observing backprojected contributions to circles centered around the object of interest.

8) 1 The Shah function III infinite sequence of Dirac impulses separated and Δ is defined as an ∑ weighted with a distance Δ. Therefore. 2Δ2 4Δ2 t t (2. Because of its shape in the Fourier domain. Without any filtering of projection data..4. this filter is called a rampfilter. In a practical implementation. IIIΔ (t) Δ ∞ k=−∞ δ(t − kΔ).3: Illustration of projection followed by backprojection. one way to avoid discontinuities in this t periodic Fourier transform is to multiply Fg∞ with a rectangle of width Δ−1 and t sample its inverse Fourier transform. and sin(πx) . i. gs [k] = g(kΔt ) = ⎪ ⎩ − 2 1 2 k odd 2Δ k π sinc(x) t (2. From the theory of sampled signals (see for instance Bracewell [5]). As shown in Fig. 2. all signals involved are digital. . this operation has a low-pass filtering effect on the image (see text). Hence. the rampfilter g∞ must be bandlimited and sampled before it can be used. the Fourier transform of the sampled rampfilter is Δ−1 -periodic.12 Analytical reconstruction methods            Figure 2. πx Sampling of g with sampling distance Δt results in the sequence ⎧ 1 ⎪ 4Δ2 k=0 ⎨ t 0 k even . Here. we know that sampling with a distance Δt is equivalent to convolution with the Shah function1 IIIΔ−1 t in the Fourier domain. This inverse Fourier transform is given by g(t) = F −1 = Δ−1 t Π(ρ/Δ−1 ) − Λ(2ρ/Δ−1 ) t t 2 1 1 sinc(t/Δt ) − sinc2 (t/(2Δt )).6) where Π and Λ are rectangle and triangle functions respectively. g∞ is a generalized function satisfying Fg∞ = |ρ|.e.7) (2.

For each angle θl .4: By bandlimiting the rampfilter so that no frequencies higher than Δ−1 /2 exist.2 Weighted filtered backprojection The Weighted Filtered Backprojection method (WFBP) by Stierstorfer et al.9) can be made finite by truncating gs [k] to the double length of this interval. . resulting in frec (x. θ] is non-zero.. [100]. Therefore. i. Nt −1 q(t. (2. a continuous projection q(t. tNt −1 .e. is an approximative reconstruction method for helical cone-beam CT. θNθ −1 .2 Weighted filtered backprojection 13      Figure 2. y) = π Nθ Nθ −1 q(x sin θl − y cos θl . and one of the main components in iterative methods studied in later chapters. θl ) is created by means of linear interpolation. The three main steps are . Discretization of the backprojection operation is done in the following way. we give a presentation of the WFBP method.11) 2. We are only interested in the result of this convolution in the interval where p[k.2. θl ).10) The final image can now be created by approximating the integration over angles with a summation.. θ] = Δt l=−∞ gs [l]p[k − l. In the following. θl ]Λ t − tk Δt ... and radial coordinates t0 .. the Fourier transform of the bandlimited and sampled rampfilter t becomes continuous. l=0 (2. θl ) = k=0 q[k. . mainly based on [100]. Suppose that projection data have been collected for the angles θ0 . θ] (2.9) where p[k.. the summation in (2. Rampfiltered projection data can now be calculated as ∞ q[k. θ] are the original projection data for a certain angle θ.

14

Analytical reconstruction methods

Physical detector

Virtual detectors

y x

y x

z x

z x

Figure 2.5: Left: the original cone-beam geometry. Right: the semi-parallel geometry obtained after rebinning. Because of the specific shape formed by the rays, this geometry is also called the wedge geometry [111]. 1) Rebinning of cone-beam data to semi-parallel data (Fig. 2.5), 2) Ramp-filtering of rebinned projection data, and 3) Normalized backprojection. To understand these steps, the geometries of Fig. 2.6 and 2.7 must be understood. Given a certain source angle α, the location of the source is determined by xS (α) yS (α) zS (α) = = = RS cos α, RS sin α, and Pα , 2π (2.12)

2.2 Weighted filtered backprojection

15   

Figure 2.6: Scanning geometry in the xy-plane.   

Figure 2.7: Scanning geometry in the yz-plane. where RS is the distance between the source and the isocenter, and P is the patient table feed, or equivalently, the pitch of the helix. An arbitrary ray is determined by the angles α, β and a continuous variable q ∈ [−1, 1] describing the slope of the ray, where q = ±1 corresponds to rays hitting the top and bottom of the detector respectively (see Fig. 2.7). Given these parameters, arbitrary points along the ray can be calculated by xα,β,q (l) yα,β,q (l) zα,β,q (l) = = = xS (α) − l cos(α + β), yS (α) − l sin(α + β), and zS (α) + ql tan κmax . (2.13)

Here, l determines a particular point on the ray. For instance, l = 0 points out the source position, and l = RS + RD points out the position of the detector element.

16

Analytical reconstruction methods

Fig. 2.6 shows how an arbitrary ray is uniquely determined in the xy-plane, either by specifying α and β or by specifying the angle θ and the orthogonal distance t to the isocenter. Given α and β, the parameters θ and t can be calculated from θ = α + β, t = RS sin β. (2.14)

The other way around, if θ and t are known, α and β can be obtained by using the relation α = θ − β, β = sin−1 t RS . (2.15)

The procedure of translating 2D projection data parametrized with (α, β) to parallel data (θ, t) is referred to as rebinning. In the first step of the WFBP method, this is done independently for each detector row. The resulting geometry is illustrated in the right image in Fig. 2.5. Projected onto the xy-plane, the rebinned projections look perfectly parallel. However, the rays diverge in the z-direction and the rebinned projections are therefore called semi-parallel projections. As a second step in the WFBP method, rampfiltered projections pconv (θ, t, q) are obtained from p(θ, t, q) by applying the rampfilter from the previous section. Similarly to the rebinning, rampfiltering is done on each detector row independently, i.e., data are not mixed between detector rows. The rampfilter can be combined with a low-pass filter m(t) to shape the frequency response of the CT system. Commercial systems are shipped with a variety of such filters, designed for different clinical applications. Throughout this thesis, we refer to these filters as resolution modifying filters. Similarly to two-dimensional backprojection, the WFBP backprojection constructs a voxel volume as a sum of partial backprojections
Nπ −1

frec (x, y, z) =
n=0

fθn (x, y, z) ˜

(2.16)

˜ over different angles θn , where Nπ is the number of projection angles per half turn. However, in contrast to (2.11), each term in this equation contain contributions ˜ from several θn -angles, namely θn = θn + kπ, k = 0, 1, .... The reason for dividing the backprojection in such partial backprojections is illustrated in Fig. 2.8. This figure shows how a certain voxel may receive one, two or three contributions from ˜ one θn -angle depending on the location of the voxel. Because of this, and the weighting described in the previous paragraph, normalization must be performed ˜ for each individual θn -angle. Below, we describe how the contribution from one such angle is calculated. A more detailed description of the backprojection is given in Section 10.5.1. Given a point (x, y, z) in which we want to know the backprojected values, ˆ ˆ a first step is to calculate the projection (tk , qk ) of this point onto the virtual

y. To alleviate these artifacts.10 for different choices of Q. This is given by ˆ tk qk ˆ = = ˜ ˜ x sin(θn + kπ) − y cos(θn + kπ).18) ˆ Knowing tk and qk . we see that the artifacts caused by the sharp transitions are almost perfectly eliminated.21) This function is illustrated in Fig. . In Fig. or more contributions. z) = ˜ 1 WQ (ˆk ) q k ˜ ˆ ˆ WQ (ˆk )p(θn + kπ.9 a). showing a WFBP reconstruction with Q = 0. detector plane illustrated in Fig. 2. tk . voxels receive two and three contributions respectively. y. q k (2. (2.8: Example of how different voxels receive different numbers of contri˜ butions during backprojection from a certain angle θ.9 b). (2.2 Weighted filtered backprojection 17          Figure 2. qk ). qk ). ˜ z − P θn + kπ − sin−1 ˆ tk RS /2π 2 ˜ ˜ ˆk RS − t2 − x cos(θn + kπ) − y sin(θn + kπ) × RS . tk .19) is modified into fθn (x. two. to t ˆ As shown in Fig.17) (2. Voxels located in the lightly shaded area receive a contribution from one focus position.20) where the weighting function WQ (q) is given by given by 1 WQ (q) = cos π |q|−Q 2 1−Q 2 |q| < Q Q ≤ |q| < 1 . z) = p(θn + kπ. tk . 2. the sharp transitions between regions receiving one. and Nk is the number of non-zero contributions in the sum over k.5. give rise to severe non-local artifacts. qk ) is obtained by linear interpolation with respect ˆk and qk . 2.19) ˜ Nk k ˜ ˆ ˆ where the value of p(θn +kπ. In the medium and dark shaded areas.2. a weighted and normalized partial contribution can be calcuˆ lated as 1 ˜ ˆ ˆ fθn (x. the backprojection formula in (2. h/2 (2.7. 2.

3. 1) 0.9: WFBP reconstructions of the Turbell Clock phantom [110] with Q = 1.0 b) Q = 0. Q = 0.18 Analytical reconstruction methods a) Q = 1. . 0.4 0.0 and Q = 0.4 Normalized row coordinate q ∈ (−1. Greyscale window ±20HU.0 0.2 0 0.8 1 Figure 2.8 −0.10: The weighting function WQ (q) for Q equal to 0.0 1 Q = 0.7 Figure 2.6 0.8 Weighting 0.0. and 1.3 Q = 0.6 0.0.7 respectively.2 0 −1 −0.2 0.6 −0. 0.7 Q = 1.7.4 −0.

1) and is illustrated in Fig. 3. k refers T to a certain iteration and each component fj in fk = f1 · · · fN represents a certain voxel (see (3. Thus. the projection of the estimate.1. iterative methods generate a sequence of estimates that eventually should converge towards a good solution. By keeping αk constant with respect to k and letting Δ = I − αPT P. Application overview. originally developed for iteratively solving Fredholm equations of the first kind [55].Chapter 3 Iterative reconstruction methods 3. Here.8)). The matrix P ∈ RM ×N represents a linear model of the physical acquisition process. .2) = Δk+1 f0 + (Δk + · · · + Δ + I)αPT pin . and let pin ∈ RM represent input projection data. The updating formula for the Landweber method is given by fk+1 = fk − αk PT (Pfk − pin ) (3. How this works in principle is illustrated here with the Landweber method. Let the estimates of linear x-ray attenuation coefficients be represented by vectors fk ∈ RN . The difference Pfk − pin is then used to create a new estimate by backprojection with the adjoint operator PT ∈ RN ×M .1 A basic example. A detailed discussion of the relation between the vector fk and corresponding image by means of basis functions is found in the following section. P maps the estimate fk onto Pfk . the result of which is multiplied with the step length αk . we see that fk+1 ··· = = Δfk + αPT pin Δ2 fk−1 + (Δ + I)αPT pin (3. and finally subtracted from the old estimate. As mentioned in Chapter 1.

5) (3. a fact which has forced the CTindustry to rely on non-exact methods. then the first term become zero and the expression (Δk + · · · + Δ + I) converges to (I − Δ)−1 . Compared to analytical reconstruction. projection data Pfk are first calculated. In contrast. This difference image is finally multiplied with a step length α and is added to the old estimate fk . let PT (Pf − pin ) = ∇z(f ). However. These are then subtracted from the input data and a difference image (gradient) is obtained by backprojection with the adjoint operator P. 2 2 (3. Since z(f ) can be identified as z(f ) = 1 Pf − pin 2 . there are cases where this drawback is outweighed by the relative ease by which the iterative methods handle the modeling problem. In order to calculate a new estimate fk+1 . Iterative methods can simultaneously produce accurate images and make use of all available data.1: Illustration of the Landweber method. Some examples where iterative methods have been found worthwhile to explore are listed below.1) is recognized as a steepest descent method for minimizing z(f ) [71]. if PT P is positive definite and α is chosen so that the spectral radius of Δ is less than one. This leads us to the fixed point f∞ = lim fk = (I − Δ)−1 αPT pin = (PT P)−1 PT pin .4) the method (3.20 Iterative reconstruction methods     Figure 3. In helical cone-beam CT. Any analytical method faces the problem of modeling the inverse of every physical phenomenon involved in the data capture. there is no exact method able to exploit redundant data for better Signal-to-Noise Ratio (SNR). k→∞ (3. at this time of writing and to the best of our knowledge.3) In the updating formula (3. Therefore. most iterative methods are slow. . • Redundant data. iterative methods require an emphasis primarily on modeling the forward versions of all such processes.1).

3.2 Different types of iterative methods

21

• Better noise models. Recently, several authors have shown how improved statistical models of input data can improve the SNR in helical cone-beam CT [73, 16, 121, 127, 109, 98]. • Incomplete data. One example of how to suppress artifacts due to missing data in circular cone-beam CT is given by Zeng et al. [123]. Here the Feldkamp method is applied iteratively to reduce errors due to the incomplete circular source trajectory used for data capture. • Resolution recovery. In studies by Nuyts et al. [73] and Thibault et al. [109], iterative methods were shown to increase the spatial resolution in comparison to filtered backprojection methods. • Beam hardening correction. De Man et al. [21] and Elbakri and Fessler [23] have shown that by simulating multi energetic x-rays in the acquisition model, beam hardening artifacts can be significantly reduced. • Metal artifact reduction. De Man et al. [20] have shown that iterative methods can be used to reduce artifacts caused by the presence of strongly attenuating objects. In the following sections, we give a brief presentation of different types of iterative reconstruction methods, as well as operations and techniques that frequently appear in iterative reconstruction methods.

3.2

Different types of iterative methods

Most iterative reconstruction methods are based on the minimization or maximization of a cost function z(f , pin ). Such a cost function is typically expressed as z(f , pin ) = zD (f , pin ) + zR (f ), (3.6)

where zD measures the difference between the projected volume Pf and input data pin , and zR is a regularization term enforcing some kind of smoothness on the reconstruction result f . If the cost function is convex, any globally convergent optimization method will reach the optimum. However, from a practical point of view, the choice of optimization method is critical, since iterations are often stopped before the fixed point has been reached. Xu et al. [117] have divided iterative methods into three categories: conventional algebraic reconstruction techniques, iterative statistical methods, and IFBP methods. Below follows a short description of each of these categories. Conventional algebraic reconstruction techniques include Algebraic Reconstruction Technique (ART) [31], Simultaneous Algebraic Reconstruction Technique (SART) [1], Simultaneous Iterative Reconstruction Technique (SIRT) [30]

22

Iterative reconstruction methods

and the Landweber method presented above, among others. It was shown by Jiang and Wang [40] that under certain circumstances, all of them minimize weighted square norms similar to (3.5). Statistical image reconstruction methods are based on the Poisson model for the photon intensity. This is used for deriving the conditional likelihood P (pin |f ) for the acquired input data given a certain vector f of linear attenuation coefficients. Bayes’ rule gives an expression for the reciprocal conditional likelihood P (f |pin ) = P (pin |f )P (f ) . P (pin ) (3.7)

The aim of statistical methods is to maximize log P (f |pin ). If P (f ) is assumed to be constant, the method is called a Maximum Likelihood (ML) method. Otherwise, the method is called a Maximum a Posteriori (MAP) method. Many methods for maximizing the ML and MAP functions exist. Examples of these are Maximum Likelihood Expectation Maximization (MLEM) [56], the convex method [57], and the ordered subsets convex (OSC) method [47]. The ML and MAP methods are known to produce images with better signal to noise ratio, and many recent developments toward faster methods make these methods promising. Recently, interesting results have also been presented from using the simplified cost function obtained by approximating log P (pin |f ) with its second order Taylor expansion [50, 109]. Iterative Filtered Backprojection (IFBP) methods are based on iterative application of FBP methods. The simplest IFBP method is obtained by substituting PT with Q in the Landweber method, where Q is a FBP method. Since the FBP method is used in each iteration, certain artifacts caused by non-exactness of Q are very rapidly suppressed. Therefore, for the purpose of suppressing such artifacts, IFBP methods are usually much faster than other iterative methods. For attenuation correction in Single Photon Emission Computed Tomography (SPECT), IFBP was proposed by Chang [8], and Walters et al. [113], while Medoff et al. [66] used IFBP for improved reconstruction from limited angles. Censor et al. [7] presented and analyzed the general framework of Iterative Data Refinement (IDR), that includes IFBP as a special case. It was later shown by Xu et al. [117] that for emission tomography, IFBP methods have computational advantages over the conventional and statistical methods. Furthermore, the SNR in relation to Contrast Recovery Coefficients (CRC) of IFBP methods were comparable to those of regular FBP. Similar results for transmission tomography were later presented by Nuyts et al. [73]. Recently, Zeng et al. [123] employed an IFBP method for reduction of circular trajectory artifacts in the Feldkamp algorithm. Further acceleration of IFBP methods have been suggested by Riddell et al. [83], and Delaney and Bresler [22]. IFBP methods, which are the main topic of this thesis, will be presented in more detail in Chapter 5, where we introduce the Regularized Iterative Weighted Filtered Backprojection (RIWFBP) method.

3.3 The projection operator

23   

    

! "    

  

Figure 3.2: Illustration of how a two-dimensional array of image coefficients are mapped onto a continuous function by interpolation with a bilinear basis function. The sampling distance in the y-direction is assumed to be equal to the sampling distance in the x-direction.

3.3
3.3.1

The projection operator
Matrix representation

The following derivation of a matrix formulation of the projection operation P has mainly been inspired by a paper on local basis functions by Hanson and Wecksung [33]. A continuous three-dimensional image, i.e., a not necessarily continuous function that maps R3 into R, can be represented as a linear combination of basis functions as
N

fc (r) =
j=1

fj bj (r),

(3.8)

where r ∈ R3 is a coordinate vector and bj : R3 → R are the basis functions (for the 2D case, see Fig. 3.2). Usually, these basis functions are translated copies of a single basis function b(r) so that (3.8) becomes
N

fc (r) =
j=1

fj b(r − rj ).

(3.9)

24   Iterative reconstruction methods               Figure 3. In the three-dimensional case. where each rj corresponds to the center of a certain voxel. The result is then integrated to get the final result pi = R2 wi (r)fc (r)dr.11) T From this expression. The image fc (r) is first multiplied with an irradiation function specifying the photon flux in different areas of the image. the contribution from a continuous three-dimensional image fc : R3 → R to a certain detector reading i can be calculated as pi = wi (r)fc (r)dr. They constitute a model for the source and detector system and contain values of the photon flux for a certain reading i in a certain point r. Fig. Insertion of (3. if nonlinear effects are neglected. (3.10) R3 The three-dimensional functions wi (r) will throughout this thesis be called irradiation functions. The simplest and most common choice of functions wi are Dirac lines.9) into (3.3: Illustration of a linear model for projection generation in the case of two dimensions. we see that the projection data p = p1 can .3 shows a linear model for the acquisition process in the case of twodimensional functions. 3. More elaborate functions such as strips or sums of several Dirac lines can be used to create more realistic models of the acquisition process (to be discussed in Chapter 9). resulting in pure line integral contributions.10) yields N N pi = R3 wi (r) j=1 fj b(r − rj )dr = j=1 fj R3 wi (r)b(r − rj )dr ··· pM (3.

Later in this section.12) yields a line integral taken through the basis function.14) . if wi (r) is a Dirac line. (3. Below. the above computation can be efficiently implemented as a simple lookup in a table of precomputed values (see for instance [124] and [69]). This observation takes us to the important notion of footprints. Ohyama et al. t) of the basis function b(r) is defined by g(s.13) According to the Fourier slice theorem. (ii) allow for cost-effective implementation of forward projection and backprojection operations.3 The projection operator be calculated by a matrix multiplication p = Pf where f = f1 the components of P ∈ RM ×N are given by pij = R3 25 ··· fN T and wi (r)b(r − rj )dr. t) = R b(t + ls)dl.3. Assuming that the sampling distance equals Δx . [76] and Danielsson and Magnusson [12] solved the verification problem by studying the Fourier transforms of the involved functions. Then the footprint g(s. Requirement (i) is easy to understand and verify for basis functions which are separable as b(x. and (iii) contain a minimal amount of frequency components higher than the Nyquist frequency.2 Basis functions The choice of basis function b(r) as well as the irradiation functions wi (r) affect the result of an iterative method. the Fourier transform of the repeated basis functions is given by F 1 IIIΔx ∗b Δx (u) = (IIIΔ−1 ·Fb)(u). t) with respect to t equals the Fourier transform of b(r) on s⊥ . Rotationally symmetric o basis functions and basis functions implicitly defined by their footprints are harder. Let s ∈ S 2 ⊂ R3 be a direction vector and let t ∈ s⊥ ⊂ R3 describe a displacement in the orthogonal complement of s. the Fourier transform of g(s. (3. and the K¨hler basis function (trilinear interpolation) [51].3. y) = bx (x)by (y). since these give rise to aliasing artifacts. the footprint can be used to study Fourier domain properties of a basis function. (3. x (3. footprint tables can be used for efficient implementations of P [69]. Thus. Examples of such basis functions satisfying this requirement are the Siddon basis function (nearest neighbor interpolation) [92]. examples will be given of basis functions that are defined by specifying their footprints rather than directly specifying b(r). In cases where the footprint is invariant to changes in direction s and the irradiation functions wi (r) are Dirac lines. 3.12) For simple irradiation functions such as Dirac lines and strips. Clearly. A good basis function should (i) be able to accurately represent a constant function. we discuss the first of these properties in more detail.

e. 3. but also false high frequency components.26 ∞ Iterative reconstruction methods where IIIΔx (t) = Δx k=−∞ δ(t − kΔx ) (see p. (3. 61]. During projection generation. Thus. 0 . A simple example of a basis function that violates this requirement is illustrated in Fig. v) = Δ2 sinc2 (Δx x u2 + v 2 ). However. such a relation between the detector and voxel density may give rise to aliasing. it is clear that there are non-zero contributions at the impulse points of IIIΔ−1 . 12). which occurs as a special case of a projection method proposed by Peters [78].12) into a line integral along the line corresponding to measurement i. 1) In order to capture all information delivered by the input data. Otherwise. 3.15) The Fourier transform of this basis function is given by (Fb)(u. they are close to zero outside the Nyquist frequency of the reconstruction volume.3.16) and is exactly zero in the impulse points of IIIΔ−1 along a line aligned to one of the x main axes.. This aliasing of neighboring DC-components was named “DC-aliasing” by Danielsson and Magnusson [12]. it is desirable to use a relatively high voxel density in comparison to the ray density. There are at least two reasons to consider other types of irradiation functions. To be able to represent the underlying constant function. This basis function. otherwise (3. Its Fourier transform Fb must be close to zero in all points where the impulses of IIIΔ−1 are located.5. t) = Δx − |t| . This method is further discussed in Section 3. This is realized by first studying the continuous . |t| ≤ Δx . the basis function (to be convolved with the samples) must fulfill the following criterion.4. false frequency components appear when x this basis function is used for approximating a constant function. is defined by its rotationally symmetric footprint g(s. i. Another rotationally symmetric basis function based on the SinCot filter by Magnusson [64] has been presented and evaluated in [12] and [101].4 and 3. not only a DC-component will appear in the result.3. One well-known example of rotationally symmetric basis functions satisfying (i) and (ii) above is the class of generalized Kaiser-Bessel basis functions (also known as blobs) proposed by Lewitt [60. The forward projection method by Joseph [41] overcomes this problem by scaling the footprint depending on the projection angle. except in the x origin. if the Fourier transform is considered along a line tilted 45◦ away from one of the main axes. These functions are effectively bandlimited.3 Irradiation functions The most common type of irradiation function is a Dirac line transforming the volume integral in (3.

this type of aliasing will occur for a wide range of sampling densities Δz .4: Illustration of the periodic Fourier transform of a sampled image and √ zero crossings for the function Δ2 sinc2 (Δx u2 + v 2 ). The next step is to see whether the assumptions of the sampling theorem [5] are violated during sampling of this continuous projection. Therefore.3. x projection of the continuous image obtained by convolution by the basis function. aliasing may occur during projection. 3. 2) The line integral model of the acquisition process is rather crude since the . If the x-ray detector is configured so that the rays are relatively sparse in comparison to the voxel density. such irradiation functions can be used for suppressing aliasing in the projection generation process.3 The projection operator 27        Figure 3. As illustrated in Fig. for diverging rays appearing in the z-direction after semi-parallel rebinning in WFBP. Using an irradiation function that consists either of several Dirac lines or a strip is similar to convolving the continuous projection with a certain lowpass filter prior to sampling.6.

.28 Iterative reconstruction methods                   Figure 3. −Δ−1 . . .... the repeated DCx x components are perfectly eliminated for 0◦ . However.5: Two intersections through the Fourier transform shown in Fig.. 0. Δ−1 . √ Since Fb is zero for ρ = u2 + v 2 = . 3. this is not true in the 45◦ direction where false high frequency components are introduced..4.

These widths were chosen to be the image . High frequencies located close to the focus are projected to relatively low frequencies on the detector. Ziegler et al. strip integrals (see Fig. Later. showed that suppression of these artifacts can be done by using divergent strip irradiation functions in combination with blob basis functions [124]. and Ohyama et al. gantry rotation and sizes of focus and detector elements are neglected.3 The projection operator 29                  Figure 3. A better model would take into account the smoothing caused by these elements. De Man and Basu [17] proposed a technique for generating projections that resembles the Joseph interpolation but instead of a triangular interpolation function their interpolation employs two convolved rectangle functions with different widths. 3. In the papers by Hanson and Wecksung [33]. [76].6: An illustration of how aliasing occurs during projection generation. The problem of aliasing was pointed out by Mueller et al. while high frequencies located far from the focus are projected to relatively high frequencies on the detector. [68] and a space variant scaling of the basis functions (in this case blobs) was suggested to suppress aliasing artifacts. Various examples of irradiation functions other than Dirac lines exist in the literature. better results were achieved with strips than with Dirac lines.3.7) were used in combination with the square basis function in two dimensions. In both papers.

This method has the advantages of being simple to implement. the first rectangle would correspond to the basis function and the second to the irradiation function. Danielsson and Mag- .8. The first step in the Joseph method is to divide the image into a set of lines {Li }Nx along the coordinate axis that is maximally i=1 orthogonal to the line Γ. and producing less aliasing artifacts than the method proposed by Siddon [92]. computationally efficient. Finally. we give a brief description of the method in two dimensions. sampling distance and the spatial dependent ray distance respectively. 3. We believe that this idea should be highly beneficial to apply in the divergent projection field in Fig. An extension to three dimensions is straightforward but more difficult to illustrate. 3.6. in terms of basis and irradiation functions. Consider Fig. 3. In Chapter 9.7: Examples of different irradiation functions.30 Iterative reconstruction methods  Figure 3.3. in the subsequent implementations and experiments we have used the forward projection method proposed by Joseph [41].7) for anti-aliasing and more accurate modeling of the acquisition process is investigated. Suppose that the line integral through an image along a line Γ is to be calculated.4 The Joseph forward projection method Unless otherwise noted. i=1 The Joseph method does not suffer from DC-aliasing. Here. The next step is to use linear interpolation to calculate contributions from each intersection between Γ and the lines {Li }Nx . 3. Thus. a i=1 length correction is made by multiplying the sum of contributions with Δx / cos(γ). where γ is the smallest angle between Γ and the coordinate axis perpendicular to the lines {Li }Nx . the effect of using several Dirac lines per reading (see Fig.

the sum of contributions is multiplied with Δx / cos(γ). the Fourier transform is stretched with a factor 2. 3. Depending on the slope of the line integral to be calculated. Except for the compression factor cos(γ(s)). the image is divided into a set of lines such that the angle γ does not exceed 45◦ .8: Illustration of the Joseph projection method. a contribution is calculated by means of linear interpolation. Theorem 4. By reasoning in the same way as was previ1 ously done for (3.9 [110].15).3.15). 1 This fact was pointed out to the author by Dr. Michel Defrise. all non-central repetitions of the DC-component are eliminated also in the 45◦ -directions. One problem with the Joseph method is that the footprint violates the Helgason Ludwig conditions (see [70].8. At each intersection between the line and this set of lines. 3. 13] showed this by studying the footprint 1 cos(γ(s)) g(s.17) where γ(s) is the angle in Fig. this expression is identical to (3. This results √x in a triangular footprint g(s. it can be useful to imagine a trilinear basis function sheared as illustrated in Fig. Therefore. we see that since the footprint is compressed with a factor √2 √ in the 45◦ -direction. As a final step. t) = Δx − |t| cos(γ(s)) 0 |t| ≤ Δx cos(γ(s)) otherwise (3. In cases where a basis function interpretation of the Joseph method is needed. nusson [12. Therefore. . and in the directions arctan(±1/2) as well. t) with a width that varies between 2Δx and 2Δ2 .3 The projection operator 31                 Figure 3.1)1 . there is no well-behaved basis function (possible to approximate with a function in the Schwartz space) that corresponds to this footprint.

if white noise is added to the projection data. In the context of tomographic reconstruction. Naively applying the Landweber method to find the pseudo inverse of P as described in Section 3.32 Iterative reconstruction methods      Figure 3. In [60].. Therefore. The space of functions that maps the real line to itself is too large and contains many solutions that we would consider incorrect or noisy. Therefore. Lewitt propose to use rotationally symmetric generalized Kaiser-Bessel basis functions (also known as “blobs”) for iterative reconstruction. 3. This is caused by (i) ill-posedness of the tomography problem2 and (ii) insufficient restrictions on the space of possible solutions.1 leads to strong amplification of these noisy components and unusable reconstruction results. The problem with insufficient restrictions on the space of possible solutions is illustrated in Fig. In the context of image reconstruction. i. if a small perturbation of input data may lead to a large perturbation of the solution [32]. constraints must be added to restrict the solution space to certain well-behaved functions. A presentation and discussion of the general problem of ill-posedness and the Fredholm equation of the first kind is found in [32].4 Regularization A common problem with iterative image reconstruction methods is that the iterates tend to become distorted by high frequency artifacts and noise as the number of iterations grows. or if the solution does not continuously depend on the input data. projection data pin contain only small amounts of high frequencies.e. ill-posedness means that the projection operator Pphys suppresses high frequencies rather efficiently.9: Illustration of the Joseph basis function for different angles.10.e. . The figure is based on a similar figure in [110].. The amplitude 2 A problem is said to be ill-posed if its solution is not unique. This is called regularization. high frequency components of the projection data will quickly be dominated by noise. 3. at least four different regularization techniques appear in the literature. i.

[126]. [96] proposed the use of sieves to suppress the enhancement of noise as the number of iteration is increased.4 Regularization 33     Figure 3. Another.3. sieve regularization would be identical to regularization with a smooth basis function. spectrum of these functions is very close to zero outside the Nyquist frequency of the image. K > 0. The practical problem with this method is to choose a good stopping criterion. Snyder and Miller [95]. f can be identified as a sequence of weighted Dirac impulses corresponding to the coefficients of the f vector. As a final step (after convergence). For instance. the basis function is used for controlling the frequency content of the reconstructed image.10: A simple example of fitting a function f : R → R to a number of measurements.18) where K(r) is called the sieve kernel and is usually a smooth function. and Snyder et al. Note that the continuous formulation (3. Stopping an iterative method before convergence is theoretically unsatisfactory. perhaps more common interpretation is to design the iterative method so that the resulting coefficient vector f lie in the sieve Sdiscrete = f ∈ RN : f = Kf . Then. K > 0. f is convolved with the sieve kernel to get f . f ∈ RN + (3. for instance a Gaussian. In this way. The iterative method is designed to estimate f by introducing a convolution with the sieve kernel before projection as well as after backprojection in each iteration. Another popular method for avoiding enhancement of noise is to stop the method before the fixed point has been reached (see for instance Veklerov and Llacer [112] and Hebert et al. and K can be identified as the basis functions. The filled circles illustrate the measured points and the curves illustrate two possible functions f1 and f2 that interpolate the measured data. This kind of regularization has successfully been applied to iterative image reconstruction in CT by Ziegler et al.19) where K correspond to a convolution with a discrete sieve kernel. since this means that a sub-optimal point is preferred over the optimal solution .18) leaves room for principally different implementations. [34]). This means that the iterative method is designed so that the result will lie in the so called sieve S= f (r) : f (r) = R3 K(r − r )f (r )dr . f > 0 (3. This choice must be done with respect to the particular optimization method being used.

This term is designed so that certain structures and properties considered likely to appear are penalized less than other less likely structures.6)). or equivalently. Regularization with this type of penalty functions is a major topic of this thesis.20) where β is called regularization parameter or hyperparameter. . of size 3 × 3 × 3. These penalty functions have the general form N N zR (f ) = β i=1 j=1 dij V (fi − fj ).21) where ∇x . and determines how different contrast levels should be penalized.34 Iterative reconstruction methods to the problem as it has been formulated mathematically. (3. Another penalty function that recently has appeared in the CT literature is the Total Variation (TV) norm N zR (f ) = β f TV =β i=1 (∇x f )2 + (∇y f )2 + (∇z f )2 . i i i (3.g. Gibbs priors [63]. y-. ∇y . The function V is called potential function. One way to improve the problem formulation is to use a penalty term zR (f ) in the cost function (the second term in (3. and dij are usually the inverse distances between voxels i and j within some small neighborhood. One widely used choice of penalty function is based on the theory of Markov Random Fields (MRF). 9]. and will be further discussed in Chapter 5. The conclusion is that the mathematical problem formulation lacks some feasibility parameter not yet accounted for. Some results have indicated that this penalty term can be useful in reconstructions from projection data that have been very sparsely sampled along the angular direction [93. e. and ∇z correspond to differentiations in the x-. and z-directions respectively. 7 and 8.

and parameters describing the scanning geometry.1. Forchheim).1). Noise was added using the procedure suggested by Fuchs [28]. 4. To make the physical model even more realistic. Let I0 be the expected number of non-attenuated photons traveling along the ray i from (4. To simulate the finite extension of the source and the detector. CTSIM/DRASIM in combination with the gantry rotation model calculates the contributions after taking logarithms as ⎛ ⎞ N 1 pi = − log ⎝ exp − fpha (r)dl ⎠ (4. and calculates a mean value of several line integrals instead of just one single line integral. this is not a problem for the present investigation. The main input consists of a mathematical phantom fpha (r).j where N = NSW × NSL × NDW × NDH × NGR . This program produces projection data in the form of line integrals. the number of attenuated photons I is approximately Poisson distributed . typically N = 35 = 243. the program divides these into NSW × NSL and NDW × NDH equidistantly sampled points respectively (see Fig. since the same input data can be used for several different experiments. Then. (4. Thus.2) As shown in Table 4.1) N j=1 Γi.Chapter 4 Simulation and evaluation 4.1 Generation of input projection data Noise-free input data for the experiments in the following chapters were generated with the CTSIM/DRASIM program package (Siemens Healthcare.1). However. which makes the computation of these simulated projection data very expensive. modeling of the gantry rotation has been added by calculating each angular measurement as the mean of NGR sub-measurements.

8 Active detector area (xy-plane) dβ 0. .77◦ a Pitch angle γP 1.9 Active detector area (z-dir) dz 1. both source and detector are divided into subunits. NDH 3 Source subsampling along xy-plane NSW Source subsampling in z-dir NSL 3 3 Angle subsampling NGR 0. It is calculated as γP = tan−1 (P/(4RS )) × 180/π.5mm Source width (xy) wS Source length (z) lS 12mm Anode angle η 7◦ Slicewidth S 1. NDW 3 Detector subsampling in row dir. Table 4.1: In order to realistically simulate the physical acquisition process.6mm Isocenter-Detector distance RD Number of channels (Quarter offset) Nch 672 48 Number of rows Nrows 1160 Number of projections per turn Nproj 3 Detector subsampling in channel dir.2mm Detector height h 57.2mm 25◦ Maximal fan angle βmax Maximal cone angle κmax 2.36 Simulation and evaluation      Figure 4. The final contribution is calculated as the mean of all sub-contributions.04◦ a This is the cone angle of a detector covering exactly the Tam-Danielsson window [110]. Source-Isocenter distance RS 595mm 490.1: Scanning parameters for experiments in following chapters.6mm Table-feed P 43.

the distribution of I is ap√ proximately Gaussian with variance k and standard deviation k. Rebinning parameters are shown in Table 4.386mm 260mm Linear Nuttalla 37 (4.2 Phantoms and error/noise measures The mathematical test objects for the studies presented in the following chapters.7. 4. Nt Nθ Δt tmax 1344 1160 0. It spans the full field of view. the deviation from the original phantom has been calculated as the Root Mean Squared Error (RMSE) σe = 1 |Ω| ((frec )i − (fphan )i )2 . σ) is a sample from the Gaussian distribution with expectation μ and standard deviation σ. 11 of [72]. artifacts that occur due to inexactness of a reconstruction method..5) . are the Turbell Clock phantom [110]. and the Forbild Head phantom [59]. 2.2 b). 4. windmill artifacts [94] and streaks are usually produced. As a first step in all experiments presented in this thesis. with E[I] = V ar[I] = I0 k. new parameters describing sampling in the new semi-parallel geometry are introduced. An illustration of the geometry and corresponding parameters is found in Chapter 2.4) I0 where N (μ.6 and 2. Thus. √ ˜ . and contains sharp edges orthogonal to the z-direction. For each reconstruction result presented in the following chapters.3) where k = exp(−pi ). (4. shown in Fig.2 a). input data are rebinned to semi-parallel geometry.4. Fig. When this number is large. The head phantom is shown in Fig. i. is useful for studying cone artifacts.2 Phantoms and error/noise measures Table 4.2: Parameters for the rebinning process. i∈Ω (4. Due to its sharp details in the inner ear region.2. Scanning parameters for experiments in following chapters are given by Table 4. In the rebinning process.1. The clock phantom.e. Number of channels Number of projections Parallel displacement per channel Maximal parallel displacement Interpolation in azimuthal rebinning Interpolation in radial rebinning a See Fig. a noisy contribution pi can be calculated as ˜ √ 1 pi = − log k + kN 0. 4.

The set Ω2 is defined as the white regions in c) and d).2 c) and d) respectively. Noise in the reconstructed images has been calculated with the formula σn = 1 |Ω| ((fnoisy )i − (fnoise−free )i )2 . different types of deviation are measured. Depending on how this set is chosen. fphan is a vector containing sampled values of the original phantom and Ω is the set of voxels taken into account for the measurement. and • Ω2 consisting only of low contrast regions. . i∈Ω (4.2: Axial images of the clock and the head phantoms.38 Simulation and evaluation a) Clock phantom b) Head phantom c) Error/noise mask for the clock phantom d) Error/noise mask for the head phantom Figure 4.6) where fnoisy represents a reconstruction from noisy data and fnoise−free represents a reconstruction from noise-free data. For the clock and head phantoms. if highcontrast edges are included in Ω. we have used two different sets: • Ω1 consisting of the complete axial image. σe will be dominated by errors caused by edge smoothing. and the summation is carried out over Ω2 . For error calculations. 4. Here. the sets Ω2 consist of the white regions in Fig. For instance.

Only simulated input data can deliver perfect ground truth for comparison. Even if the purpose of the iterative methods studied in the following chapters is to produce good results after few iterations.4. This section describes how spatial resolu- . The second data set was acquired from a head phantom using a system with 736 detector channels spanning a field of view of 500mm. differ in many ways. approximately half of the maximum pitch factor 1.2mm detector rows.3 Spatial resolution measurements 39 a) Axial section b) Coronal section Figure 4. Both CT systems were running with a pitch factor (table feed/detector height) of 0.4. and 32 × 1. due to mismatch between the model used for simulations and the physical scanning situation.4. Obviously. we have included two input data sets acquired from real medical CT systems in our tests. respectively. a necessary condition for the iterative method to converge is that this norm tends to zero as k → ∞. However.3: Thorax phantom reconstruction. it is interesting to see what the methods deliver after many iterations.2mm detector rows.g. we have calculated the Euclidean norm fk+1 − fk . 4. i. some artifacts prevalent in clinical situations cannot be captured by simulated data experiments. Therefore.3 Spatial resolution measurements Images reconstructed by analytical and iterative methods. which often can be traded for noise is the ability to preserve high frequencies throughout the reconstruction (high spatial resolution).e. and 24 × 1. The first dataset was acquired from a thorax phantom using a system with 672 detector channels spanning a field of view of 500mm. Therefore. Reconstructions of these phantoms are shown in Fig. besides error and noise measures described above. Another feature.75.3 and 4. 4. in sensitivity to noise in input data. e.

We will make a difference between in-plane (xy-plane) spatial resolution. along any direction. 1) Collection of contributing samples. the reconstructed axial image represents the axial Point Spread Functions (PSFs). which is scanned and reconstructed to a very fine grid.5). By taking the Fourier transform of the PSF. represented by Slice Sensitivity Profiles (SSPs).4: Head phantom reconstruction. Since a wire is an approximation of a Dirac line. the MTF is obtained. Let f (x) be a reconstruction of the test object. For iterative methods. MTFs and SSPs normally vary with the reconstruction volume sampling density. Below follows a step-by-step description (see Fig. assume that the direction along which we want to measure the resolution is given by the unit vector d. In a similar fashion. This is an extension of the edge MTF method originally proposed by Judy [42]. we employ the edge-based 3D MTF method by Li et al. consider the set Sγ of sample/distance . 4. which is a solid sphere with center xc and radius R. [62]. Furthermore. Therefore. Now. c) Coronal section tion has been measured in the experiments to be found in the following chapters. and cross-plane (z-direction) spatial resolution.40 Simulation and evaluation b) Sagittal section a) Axial section Figure 4. the measurements typically do not correctly describe a clinical situation. by using a very fine grid for resolution measurements as prescribed by the wire and plate methods. represented by Modulation Transfer Functions (MTFs). The most direct and common way of measuring the in-plane MTF is to use a thin wire test object directed orthogonal to the xy-plane [46]. a thin plate orthogonal to the z-axis is often used to measure the cross-plane SSP. without changing the sampling density. and has the advantage of being able to measure resolution in any location. To avoid resolution measurements on sampling densities different from the ones used for normal reconstruction.

Let the locations of these samples be defined by rj = ˜ R R +j 2 N (4. For measuring in-plane ˜ .wij >0 f (xi )wij wij .15mm. the last step is to calculate the Fourier transform of {tj }N −1 . This procedure is referred to as binning. Each element in the sequence of equidistant samples {sj }N 1 is now given by sj = i. This step reduces the set Sγ to a sequence {sj }N of equidistantly 1 located samples that represent the edge response in the direction d. From the edge response {sj }N . 1 The sphere center position xc and direction vector d determine where and along which direction the measurement is being made.4. N − 1. the derivate {tj }N −1 is 1 calculated by means of central differences as tj = sj+1 − sj−1 . 2) Binning..8) and (ii) the distances ri (xi ) satisfy R/2 < ri (xi ) < 3R/2.. N (4..9) for j = 0..11) 4) Fourier transformation.13).3 Spatial resolution measurements pairs (f (xi ). The bin size must ˜ be small enough to catch the highest reconstructible frequencies. ˜ 3) Differentiation... Two other user configurable parameters are the so called bin size Δr and the zenith angle γ. .10) i. we choose wij ˜ to depend linearly on the distance |ri − rj |. and the zenith angle must chosen large enough to ensure that each bin receives sufficiently many contributions.. we used a radius of R = 20mm (except in Fig. To obtain the MTF. (xi − xc )/ri (xi ) > cos γ 41 (4... N . . In our experiments. j = 0. ri (xi ) = xi − xc ) ∈ R × R for which (i) xi are sample points located within the cone defined by d. Note that how we describe and perform binning here is slightly different from the original presentation in [62]. While in the original presentation. . wij is a step function of |ri − rj |. an estimate of the PSF 0 is obtained by means of differentiation. Specifically. j = 1. 2R/N (4. and a zenith angle of γ = 25◦ . a bin size of Δr = 0. The set Sγ contains all information needed to proceed with our MTF measurement.wij >0 ˜ where wij = 1 − |ri − rj |/(R/N ).7) (4. 5.

[62]. However. reconstructed images of solid spheres often contain windmill artifacts.8. the proposed method is more indirect than the previously used multiple plate method.5 with a shell of thickness 0. Therefore.7 shows that there are only very small differences. 4. which could corrupt the measurements.6. Fig. resolution.1mm.5: Illustration of the edge based 3D MTF method by Li et al. Therefore. it is true that differentiating an edge response is equivalent to directly measuring the point response (except for errors introduced by the differentiation). 4. 4. this is no longer true. In linear reconstruction. this method is very similar to the 2D edge method previously used in [106]. we have compared SSP measurements using the 3D MTF method and the multiple plate method illustrated in Fig. we have modified the above method to measure surface responses by replacing the solid sphere in Fig. . for cross-plane resolution measurements. Reconstructed images of phantoms for the edge based and surface based method are shown in Fig. for nonlinear methods such as those employing non-linear regularization in Chapter 8. It seems appropriate to make the following note on edge based versus surface based methods for examining spatial resolution. To this end. Furthermore. However.42 Simulation and evaluation     Figure 4. 4. the latter being represented by the above-mentioned thin plate method. it is sometimes of interest to examine both edge responses and surface responses.

and the multiple thin plate method.8 Normalized SSP 0.2 0 −2.5 −2 −1.5 0 0. 1 Edge 3DMTF Surface 3DMTF Multiple thin plates 0.4 0.5 Figure 4.3 Spatial resolution measurements 43   Figure 4. .7: SSP measurements by the edge and surface based 3D MTF methods.6: Multiple thin plate measurement of the SSP of a system.5 Distance (mm) 1 1. By merging responses from a number of discs. whose exact positions are known.4.6 0. an oversampled SSP is obtained.5 −1 −0.5 2 2.

1mm. and the shell phantom in b) has a thickness of 0. The tests objects have a radius of 2cm.44 Simulation and evaluation a) Edge phantom b) Surface phantom Figure 4.8: Reconstructed axial sections of phantoms for edge based and surface based resolution measurements. . Greyscale window ±50HU.

we investigate the possibility to use the iterative scheme illustrated in Fig. using a projection operator P. a new improved reconstruction f1 is obtained.1. projection data are generated and subtracted from input data. In early experiments we noted that a slight lowpass filtering of input data can help to suppress overshoots in the xy direction in .Chapter 5 Regularized iterative weighted filtered backprojection 5. Much of the material here has previously been published in [106]. let pin ∈ RM denote the input data vector and let f0 ∈ RN denote a vector representing an initial voxel volume. Furthermore. the number of channels is doubled. can now be expressed by the following three steps. This step is equivalent to the rebinning step of the WFBP method. From this reconstruction. we employ the same matrix/vector notation as in Chapter 3. To simplify our presentation of finite-dimensional linear operators and linear equation systems. 5. the rebinned and preprocessed input data preb consist of 2M elements. containing cone artifacts. 5. By adding a reconstruction of these difference projection data to f0 . 2. (2) Prefiltering (preb → pmod ). Let N be the number of voxels and M be the total number of x-ray attenuation measurements. The iterative method.2). (1) Rebinning to semi-parallel geometry (pin → preb ). The basic idea is to start with a non-exact reconstruction f0 . To preserve radial resolution. which we call Regularized Iterative Weighted Filtered Backprojection (RIWFBP) is illustrated in Fig.1 Description of the reconstruction algorithm In this chapter.1 for reducing cone artifacts caused by the non-exact Weighted Filtered Backprojection (WFBP) reconstruction method [100] (see Sec. Therefore.

Nevertheless. Clearly. the artifact level around sharp z edges gets lower than for noniterative WFBP.2 for different choices of γ. and may therefore result in new artifacts. the reconstruction results [101]. [54]. In our experiments we used γ = 1/20 which corresponds to a very moderate smoothing as shown in Fig. extrapolation is done by copying the lowermost and uppermost detector rows. the matrix H ∈ R2M ×2M corresponds to a convolution with m(t) along the t direction. i. 5. The matrices Q and P correspond to WFBP method without rebinning and the forward projection operator modeling the acquisition process respectively. 5.2.1. as we will see in Sec. Given an initial image vector f0 = Qpmod .8. prefiltering in the z direction corresponds to a spatially dependent filtering of the reconstruction result. Due to the divergent beam geometry in the z direction and the extrapolation described in the previous paragraph. and a convolution with the filter kernel [γ. 2. The amplitude spectrum of this filter is illustrated in Fig. (5. along the q direction (see Fig.4. This is the main step of the RIWFBP method. f0 is set to Qpmod . this cautious choice reduces the overshoot amplitude with a factor of 5. this filter does suppress high frequencies but only with a factor of 0. (1 − 2γ).. γ]. a sequence of image vectors is generated by the update formula fk+1 = fk − αS(Q(Pfk − pin ) + βRfk ). the nonexact reconstruction . and to to suppress overshoots along the z axis.1) Here. to change the resolution properties according to the filter m(t).46 Regularized iterative weighted filtered backprojection       Figure 5. This was also theoretically motivated by Kunze et al. At the same time. 5. They also showed that a resolution modifying filter such as m(t) described in Chapter 2 can be used with IFBP methods by performing the convolution as a prefiltering step. Initially.e.1: Illustration of the RIWFBP method. Therefore. 1/3].6 and 2.7 for a description of the geometry). (3) Iterative application of the WFBP method. γ ∈ [0. To obtain boundary data for the convolution to be performed. the first step is to calculate pre-processed input data as pmod = Hpreb .

9. S correspond to the mild lowpass filter [0. they do not affect the fixed point as long as the matrix S is invertible. Unless otherwise noted.1 0. . 3. which in turn leads to faster suppression of artifacts. The matrix Q ∈ RN ×2M represents the WFBP reconstruction method without the initial rebinning step. (1 − 2γ). Iterative application of 5. 0.8 Amplification 0.5. (5.2 0 0 0. we set Δ = I − αS(QP + βR). 0.4).4 0.05] applied along all three dimensions in the reconstruction domain. Zeng and Gullberg [122]. The black curve indicate the filter used in our experiments.3) = Δk f0 + (I + Δ + · · · + Δk−1 )αSQpmod .6 0.2).2: Amplitude spectrum of the filter [γ. To find the fixed point of (5. This is the key element which distinguishes the RIWFBP from other iterative methods.4 0. and as we will see below.3 Normalized frequency 0. Throughout this chapter. The exact update step of the RIWFBP method is given by fk+1 = fk − αS(Q(Pfk − pmod ) + βRfk ). further described in the next section. The matrix R ∈ RN ×N is a highpass filter responsible for the regularization of the method. result. e. a rapid suppression of cone artifacts is obtained.2 0. 1 The described technique for deriving the fixed point was previously used by several authors.3. we employ the Joseph projector [41] (see Sec. Since this nonexact reconstruction operator is used in each update.2 yields1 fk ··· = = Δfk−1 + αSQpmod Δ2 fk−2 + (I + Δ)αSQpmod (5.1 Description of the reconstruction algorithm 1 0. P ∈ R2M ×N corresponds to a forward projector. This allows for a slightly higher step length α to be used.5 47 γ =0 γ = 1/40 γ = 1/20 γ = 1/10 γ = 1/5 Figure 5.05. The step length α ∈ R and the low-pass filter S ∈ RN ×N both correspond to operations controlling the convergence of the method. γ] along the q direction in step (1) of the RIWFBP method.2) In this equation.g.

then f∞ = f . if it exists. Thus. numerical studies must be made to investigate how aliasing and noise affect the reconstructed images. in practice. If the spectral radius of Δ is strictly smaller than one.7) . Even when the null space of QP is empty.5) The reconstruction result can be seen as an initial nonexact reconstruction with Q. Obviously. this means that (QP)−1 .6) where βxy and βz are parameters used for tuning the in-plane and cross-plane spatial resolution. the method will converge to f∞ = (QP + βR)−1 Qpmod . β = 0. the highest frequencies will be almost completely suppressed. we arrive at the expression fk = = Δk f0 + (I − Δk )(I − Δ)−1 αSQpmod Δk f0 + (I − Δk )(QP + βR)−1 Qpmod . given that f0 linearly depends on input data. 2. In turn. will amplify high frequencies very much. an undesirable enhancement of high frequencies occurs in images created by (5. (5.5). Therefore. −1]z ∗ Bx ∗ By .2 The regularization operator R We start this section by considering the case without regularization. To enable approximately independent control of the frequency responses in the xy plane and z direction. Due to the bandlimiting operation in the rampfilter and the interpolating operations in the forward projector and backprojector. fk will also depend linearly on input data. respectively.8) (5. 2. pmod is not contained in the range of P.4) Thus. −1]x ∗ By ∗ Bz + [−1. −1]y ∗ Bx ∗ Bz and Rz ∼ [−1. 2. QP will have a nonempty null space consisting of high frequency structures. we construct the matrix R as R = (βxy Rxy + βz Rz )/CN . resulting in the matrix QP + βR. respectively. If QP is invertible. The matrices Rxy and Rz are given by Rxy ∼ [−1.e. CN is a normalization constant mainly introduced for consistency with material in the following chapters. However. 5. the matrix QP works as a lowpass filter. if the reconstruction grid is finer than the detector sampling grid. The parameter β controls the amount of smoothing made by the regularization. To avoid this undesired enhancement of high frequencies and noise. a highpass filter βR is added to QP. followed by a correcting filtering with (QP+βR)−1 .48 Regularized iterative weighted filtered backprojection By expressing the geometrical sum in closed form. (5. and pmod = Pf . when β = 0. (5. (5. i.

5. and z directions. [85]. In contrast to most iterative reconstruction methods.10) i.3. we will also use the equivalent regularization operator expression N N Rf = i=1 j=1 dij (fi − fj )ei . The operator ∇s ∈ R ∇s = ∇ T x ∇T y ∇T z T .. i.12) where ei is the basis vector corresponding to one pixel. Bx . The normalization constant CN has been chosen so that the center element of the convolution kernel corresponding to (Rxy +Rz )/CN equals one. To achieve maximal rotation invariance of this Laplacian filter when βxy = βz . the parameter λ has been set to λ = 0. By . the ∼ means that multiplication with the left-hand side matrices corresponds to the convolution operations on the right-hand side.9) where g(f ) is some function measuring the deviation from Pf to input data p. and z directions respectively. respectively. These filters have the form [λ. the RIWFBP can reconstruct a LROI without any preprocessing of input data. y.5. An iterative method uses the difference between projections of reconstructed images and input data projections to calculate the next estimate.. In the following chapters. λ].11) Thus. If a LROI is used. and 1 2 3N ×N is given by 2 ∇s f 2 is a regularization term.e. . it calculates spatial derivatives in the x. (5. CN = 6 × (1 − 2λ)2 . y. and the coefficients dij are used to describe the neighborhood operation.3 Local region of interest (LROI) One weakness of iterative reconstruction methods in general is their lacking ability to reconstruct only a small local region of interest (LROI). The use of the Laplacian operator as a regularization operator can be motivated by considering the problem of minimizing the function z(f ) = g(f ) + 1 ∇s f 2 2 2 (5.e. However. in this case. (5. (1 − 2λ). This problem is illustrated in Fig.3 Local region of interest (LROI) 49 Here. s (5.093551 as suggested by Scharr et al. a large mismatch will be observed and erroneously be corrected for. using the Laplacian ∇T ∇s as regularization operator is equivalent to searchs ing for a solution that has low first order derivatives. and Bz denote onedimensional lowpass filters in the x. 5. partially or completely contained within the support of the object of interest. Differentiation of z(f ) yields the gradient descent step fk+1 = fk − α∇z(f ) = fk − α(∇g(fk ) + ∇T ∇s fk ).

i.4. 5. To understand why LROI reconstruction is possible. cone artifacts caused from objects outside the LROI are suppressed by this method. From the full ROI image. A solution to the LROI problem for iterative reconstruction methods in general has been suggested by Ziegler et al. (5.13) If Q would be exact.4. regularization and prefiltration parameters were selected so that the spatial resolution . the fixed point of the RIWFBP iteration is (see (5. 5. non-local cone artifacts from objects outside the LROI will distort the result. Initially. Thus.1 Experiments Artifacts and noise measurements Fig. 5. However. the full Region of Interest (ROI) image is reconstructed using a simple reconstruction operator Q. the embedded LROI is removed and new projection data are generated using the projection operator P. Although the LROI has been reconstructed correctly. and p LROI ( for complement) consisting of contributions from outside the ROI.3: Illustration of the LROI problem. artifacts objects outside the LROI are not suppressed.4 5. The difference between the original and new projection data is suitable for LROI reconstruction.50 Regularized iterative weighted filtered backprojection    Figure 5.5)) f∞ = (QP + βR)−1 (QpLROI + Qp LROI ). an iterative method will observe a difference and erroneously “correct” for it. Since the suggested method removes most structures outside the LROI. [127].5 b) shows an axial image of clock phantom reconstructed with the WFBP method from data generated as described in Sec. Then.e. The basic idea is to preprocess input data as shown in Fig. since Q is non-exact. 4. Qp LROI would be zero and f∞ = (QP + βR)−1 QpLROI as desired. there is a considerable difference between pfull and pLROI . In this experiment. consider partitioning of input data into pLROI consisting of contributions from the LROI.1. p = pLROI + p LROI .

Fig. A similar but less pronounced improvement can be observed between the f2 and f3 results. . Since the method has been tuned not to affect spatial resolution properties. This is followed by smaller reductions in successive iterations. After one iteration (Fig. is approximately equal for non-iterative and iterative results.4 Experiments 51       Figure 5. Further iterations result only in a small additional reduction of windmill artifacts and enhancement of high frequencies. The results observed visually in the previous paragraph are confirmed. Chapter 5). Displayed in the 40HU window.6 shows the error measures σe1 (whole image) and σe2 (without edges) for the clock phantom experiments. there is no notable change in σe1 as a function of the iteration number. the WFBP method seems to generate a mixture of two artifacts: slowly varying artifacts which we call cone artifacts. Kak and Slaney [45]. and more high-frequent artifacts commonly known as spiral or windmill artifacts. g.5. the low-frequency cone artifacts are difficult to spot and the intensity of the windmill artifacts has been reduced. we observe a reduction of σe2 from 6HU to approximately 3HU in the first iteration.4: Preprocessing of projection data for LROI reconstruction.5 c)). On the other hand. Besides aliasing artifacts known from two-dimensional reconstruction (see e. the visual appearance of the reconstructions is better. 5. 5.

f1 d) RIWFBP.0.7. Greyscale window ±20HU.5: RIWFBP reconstructions of the clock phantom. β = 0.52. βz = 1. f4 Figure 5.5. . f3 f) RIWFBP.52 Regularized iterative weighted filtered backprojection a) Sampled phantom b) WFBP reconstruction c) RIWFBP. f2 e) RIWFBP. βxy = 1. Reconstruction parameters: Q = 0.

8 shows WFBP and RIWFBP(f5 ) reconstructions of a physical thorax phantom.7 b). Fig. An interesting observation is that. 5. windmill artifacts appear in the result. many other types of artifacts exist in the WFBP reconstructions: streak artifacts.7 shows close-ups of WFBP and RIWFBP (f5 ) reconstructions of the clock phantom. 5. Other artifacts seem to be unaffected by the iterative scheme. nonlinear partial volume effects. a we employed a resolution modifying filter enhancing certain low frequencies and suppressing certain high frequencies. Comparing Fig. Pre-filtering reduces the overshoot amplitude with more than a factor of five. we observed that overshoots such as the one shown in Fig. However. 5.7 c). 5. 5. A more thorough verification of this property is found in the next section. there are no cone artifacts.7 c). this blurring seems to be fully compensated for as shown in Fig. Due to the low cone angle used in this CT system. whole image (HU) 50. we note a slight reduction of windmill artifacts when comparing the WFBP and RIWFBP reconstructions. and windmill artifacts. . In previous work. there are also cone artifacts. Prefiltering initially reduces the spatial resolution in the z-direction. the frequency response seems to be the same for the WFBP and RIWFBP reconstructions. 5.3 RMSE. Note that σe for the whole image (left plot) stays virtually constant.6: RMSE measurements σe on the whole image and on low contrast regions. Fig.6). The RIWFBP reconstructions have been made with and without the prefiltering described in (5.1 50 6 5 4 3 2 1 0 0 5 Iteration number 10 Figure 5. since the cone angle is higher than for the thorax phantom experiment.1). after five iterations. 5. 5. In the thorax phantom experiment. Fig. Again. and having found a suitable value for the regularization parameter βz from (5. without edges (HU) 0 5 Iteration number 10 50. Thus. it seems that this initial reduction of certain high frequencies protects the pre-filtered f5 -image from the overshoots in Fig. However. Similar to the experiments with the clock phantom.7 b) with 5.7 b) cannot be fully suppressed by simply increasing βz [102]. the resolution modifying filter results in predictable results not only for WFBP but also for the iterative RIWFBP method.8 c) and d). where we study modulation transfer functions of the reconstruction methods.5.4 Experiments 53 RMSE.2 50. However.9 shows WFBP and RIWFBP reconstructions of a physical head phantom. at least from visual inspection of edges Fig.

9 a). the noise drops from 15.10 shows σn measurements on reconstructions of the clock phantom as a function of iteration number. βz = 1.9 b) and d). the differences in noise level between WFBP and RIWFBP are relatively small.5. As shown in Fig. no prefilter c) RIWFBP f5 .7. βxy = 1. This is mainly because prefiltered input data are not used for the initialization. 5.8HU. 5. Greyscale window ±20HU. Initially.7: Zoomed in reconstructions five o’clock sphere close to the edge orthogonal to the z-axis. Fig. Also the windmill artifacts are reduced. β = 0. Reconstruction parameters: Q = 0. 5. but for subsequent iterations. and after ten iterations. 5.52. and as erroneously dark and bright regions in the axial image in Fig. the noise slowly increases. but this reduction is more moderate. . these artifacts are efficiently suppressed by the iterative loop.0.54 Regularized iterative weighted filtered backprojection a) WFBP reconstruction b) RIWFBP f5 . These artifacts appear as slowly varying horizontal dark and bright streaks in the sagittal image in Fig.9 c). prefilter Figure 5.3HU to 14. As the iteration number increases.

5. Greyscale window ±100HU. βz = 1. Reconstruction parameters: Q = 0.5. βxy = 1.7. .52.0.4 Experiments 55 a) WFBP reconstruction b) RIWFBP f5 c) Zoomed WFBP d) Zoomed RIWFBP f5 Figure 5. β = 0.8: Thorax phantom reconstructions.

47. . Dashed circles indicate windmill artifacts.56 Regularized iterative weighted filtered backprojection a) Sagittal image. β = 0. βxy = 1. WFBP d) Axial image. The regularization parameters were selected so that the spatial resolution of the iterative reconstruction was at least as good as for the WFBP reconstruction. Solid circles/ellipses indicate cone artifacts.9: Physical head phantom reconstructions.7. RIWFBP f5 Figure 5. Greyscale window (40 ± 75)HU.0. Reconstruction parameters: Q = 0. RIWFBP f5 c) Axial image. βz = 1.4. WFBP b) Sagittal image.

05 15 14.3 15.12. we have measured the MTF for one such filter. Therefore. (HU) 15.25 15.52. To verify that the RIWFBP in combination with resolution modifying filters produces expected results in terms of MTFs. the relative difference between the WFBP and RIWFBP f5 MTF is smaller than 2%. 5. During acquisition and rebinning. 5. and 140mm away from the isocenter. the MTF of WFBP with a resolution modifying filter should equal the product of the MTF of the WFBP without a resolution modifying filter m(t) and the Fourier transform of the filter.12. .6HU to 2.0 gives an MTF that is very similar to the WFBP MTF. 5. while βz is kept constant equal to 1. is reduced from 3. As expected.8 0 1 2 3 4 5 6 Iteration number 7 8 9 10 Figure 5. The Fourier transform of this filter is shown as the dashed line in Fig. In Fig. data corresponding to rays going through the peripheral parts of the field of view are blurred more than data corresponding to rays going through the isocenter. while βxy = 1. We see that βxy = 0.2 Spatial resolution Fig.7HU. For a translation of 140mm.0.95 14. βxy is kept constant while βz varies.0. this affects the MTF only to a very small extent. 5.2 15.0 gives a slightly sharper result than the WFBP reconstruction. 1. According to the Fourier slice theorem [70].4.1 15.13.0 gives a slightly more blurred result than the WFBP reconstruction. βxy assumes the values 0. 5.4 Experiments 57 15. and for RIWFBP with β = 0.11 shows MTFs for the WFBP reconstruction. radial MTFs are shown at the isocenter.0 respectively.15 σn .5.9 14.85 14. the frequency ρ50% (WFBP) for which the MTF of WFBP method has dropped to 50%. In the left figure.10: Noise measurements on clock phantom reconstructions.0 and 2. In the right figure. and that βxy = 2. which makes the corresponding curves overlap each other in Fig. For amplifications greater than 10%. it is interesting to see how the MTFs of WFBP and RIWFBP change as the test object is moved away from the isocenter.

f 5 . βz is constant equal to one. βxy = 0.11: MTF measurements for different choices of regularization parameters.1 0 0 1 2 3 4 5 6 Frequency (lp/cm) 7 8 9 10 WFBP. In the left image.0. For all plots. βz = 1.52.9 0. In the right image.5.0 curves overlap. Note that the WFBP and RIWFBP βxy = 1.3 0. βxy is kept constant equal to one while βz varies. f 5 .5 0.58 Regularized iterative weighted filtered backprojection WFBP RIWFBP. f 5 .12: MTF measurements in the iso-center and 140mm away from the iso-center.2 0. Regularization parameters: β = 0. while βxy varies.6 0. f 5 . f 5 .8 0. 140mm off center RIWFBP. βxy = 2. 140mm off center Figure 5. βxy = 1. βz = 0. β is equal to 0. isocenter WFBP.0 RIWFBP.7 0. f 5 .4 0.52.6 MTF 0.0 1 0.5 RIWFBP. βz = 1.8 0. f 5 . isocenter RIWFBP. The WFBP and RIWFBP curves overlap both in the isocenter and off-center cases. βz = 3. f 5 . βxy = 1.2 0 0 2 WFBP RIWFBP.6 0.0 RIWFBP.4 0.0 4 6 8 Frequency (lp/cm) 10 Figure 5. .4 0. 1 0.8 MTF MTF 0.2 0 0 2 4 6 8 Frequency (lp/cm) 10 RIWFBP.0 1 0.

5.2 0 WFBP. since no regards have been paid to the wider tails of the RIWFBP SSPs. Both the WFBP and RIWFBP MTFs follow the predicted MTF well. The deviation for low frequencies is caused by the discontinuous derivative of the predicted MTF. the SSPs do not change much in this case. βz = 3.5. the RIWFBP SSPs drop faster. In the left column. Changing βz mainly affects the amplitude of the tails. the RIWFBP SSPs are wider than the WFBP SSPs. 5. Initially. Note that the “bump” in the MTF for low frequencies gives rise to relatively slowly decaying overshoots. although the prefiltering helps to suppress overshoots.14 shows SSPs for WFBP and RIWFBP reconstructions. This product.5 and 3. When choosing a βz . since βz is held constant.13. filter Prediction WFBP. 1. mod 0 1 2 3 4 5 6 Frequency (lp/cm) 7 8 9 10 Figure 5.0 respectively.8 MTF 0. we used a test object with radius 30mm instead of 20mm. In the right column. mod RIWFBP f 5 . and for low frequencies. none of the SSPs of the RIWFBP method resemble that of the WFBP method very well. As expected.0 while βz assumes the values 0. However. βxy is held constant equal to 1.4 Experiments 59 1 0. In this particular case. which we may call predicted MTF is shown as the dash-dotted line in Fig.13: MTF measurements for the WFBP and RIWFBP method when using a resolution modifying filter. the measurement must be made over a large region. βxy is varied while βz is being held constant. except for very low amplifications. Fig. However. Unfortunately.6 0.4 0.0. Ramlak Mod. resulting in a smaller Full Width at Half Maximum (FWHM). we must take into account both the FWHM and the SSPs tail widths. in the region where the SSPs have dropped to less than 1% of the maximal values. this choice is unfair in . Considering only FWHM. a slightly larger solid sphere (radius of 30mm instead of 20mm) was required to capture this structure of the MTF.0 gives the best correspondence between the WFBP SSPs and the RIWFBP SSPs. rather than the width. To capture this structure. Therefore.

05 1 1. f 5 .8 0.60 Regularized iterative weighted filtered backprojection WFBP RIWFBP. the function log10 ( fk+1 − fk /C) is plotted against the iteration number for head phantom reconstructions. f 5 . 5.5 for the comparative studies in the previous section. For the peripheral image.8 0. the difference norm decays slowly for both the central and the peripheral image.7 × 10−4 within 30 iterations for the central image.1 0.5 −0. and one located peripherally at z = 56mm (the most peripheral image properly reconstructed by the WFBP method). βz = 0. In Fig.5 1 1.5 1 1. while βxy varies.0 0. βxy is kept constant equal to one while βz varies.1 Normalized SSP Normalized SSP 0. βxy = 0.05 0 0 −0. and over two different images: one located centrally at z = 0mm. f 5 .5 Distance (mm) 2 2.2 0 0 0. i. f 5 . βxy = 2. β = 0.4 0.0 RIWFBP. and more than 100 iterations seem to be necessary to reach . βz = 1.2 0 0 0.14: SSP measurements for different choices of regularization parameters.5 WFBP RIWFBP. All things considered.4.5 RIWFBP. this function drops down to 1.05 0. β is equal to 0.0 RIWFBP. we have chosen the regularization parameter value βz = 1. Long objects.5 2 Distance (mm) 2.6 0. In contrast. βz is constant equal to one. This has been done both for the nonregularized and regularized IWFBP. the decay is slower. For all plots.3 Higher number of iterations.5 1 0. After 100 iterations.5 Distance (mm) 2 2.6 0. In the right column.05 1 1. with regularization. βxy = 1.15. In the nonregularized case. the norm has decayed to 4 × 10−3 of its initial value. 5.5 Figure 5.0 1 Normalized SSP Normalized SSP 0.52.4 0. In the left column.5 2 Distance (mm) 2. f 5 . f 5 .0 RIWFBP. βz = 3.e. favor of the RIWFBP.

convergence. β = 0. Fig. In the following 90 iterations. with and without regularization. σe1 remains practically constant except for a small initial reduction. as a sign of reduction of cone artifacts and windmill artifacts.15: The difference norm log10 ( fk+1 − fk /C) plotted against the iteration number for one central image (z = 0mm) and one peripheral image (z = 56mm). it is small in relation to the linear attenuation values in the image.5 −1 −1. C was selected as maxk ( fk+1 − fk ). β = 0 Figure 5. Fig. we see that without regularization. 5. the measurements practically remain constant after first ten iterations. Now turning our attention to the σe2 measure. this measure initially drops from 102HU to 95HU during the first ten iterations.4 Experiments 61 0 −0. With regularization.15. log 10( f k+1 − fk /C ) . 5.5HU. β = 0 Peripheral image. 5. it slowly increases with a total of approximately 0. since acceptable results are produced within 5 iterations. except for the initial “bump” observable in all curves except σe1 with β = 0. These measurements are consistent with those in Fig. This seem to be an error caused by an initial overcompensation for differences between Pf and preb near the ends of available input data. the σe2 measure drops from 10.15.17 shows σe1 and σe2 for the “peripheral image” in Fig.3HU to 8.5 −2 −2.5. Although the initial increase in σe2 error is large in relation to the existing errors. noticeable changes occur even after 100 iterations.15 in the sense that with regularization. 5.17.5 −3 −3. we note that without regularization. These plots are very similar to those in Fig. while without regularization. 5. there is an initial drop followed by a steady increase of the error. Starting with σe1 . and remains constant on this level. β = 0.7HU in the first ten iterations.16 shows the error measures σe1 (whole image) and σe2 (low contrast regions) for the “central image” studied in Fig. and hardly visible in a 40HU window. Note that this is not a practical problem.5 −4 0 20 40 60 Iteration number 80 100 Central image. 5. For each curve.52 Central image. With regularization.52 Peripheral image.

low contrast regions (HU) 10. 5.5 0 β=0 β = 0.62 102 Regularized iterative weighted filtered backprojection RMSE (σe2 ). However. whole image (HU) 73 72 71 70 RMSE (σe2 ). low contrast regions (HU) β = 0. the regularized results very much look like smoothed versions of the non-regularized results.5 10 9.52 100 98 96 94 β=0 β=0 0 50 Iteration number 100 β = 0.52 2 1. 5. the behavior of the RIWFBP and the WFBP methods are quite different. At this point.5 9 RMSE (σe1 ).18 shows σe1 and σe2 for an image that cannot be reconstructed with the WFBP because of missing data.52 8. In the peripheral regions where the long object problem might show up. with and without regularization. when the number of iterations increases further.5 0 50 Iteration number 100 Figure 5. Fig.52 β=0 69 0 50 Iteration number 100 0 50 Iteration number 100 Figure 5. Fig. In this case.16: RMSE (σe1 and σe2 ) for head phantom reconstructions at a central image (z = 0mm).19 shows sagittal images of Head phantom reconstructions for different number of iterations. The dashed lines indicate the points where the WFBP reconstructions lose intensity due to insufficient data. we notice that the nonregularized results continue to change.17: RMSE (σe1 and σe2 ) for head phantom reconstructions at a peripheral image properly reconstructible with WFBP(z = 56mm).5 1 0. both error values are clearly reduced by the RIWFBP method. we see that most of the improvement take place during the first five iterations. Except from the initial slight increase of artifacts in images close to regions that are not . RMSE (σe1 ). while the regularized remain practically unchanged as the number of iterations goes from five to fifty. whole image (HU) β = 0. Focusing on the central regions of the phantom.

whole image (HU) β = 0.21 b).5. as expected.4. In this figure. we see that after five iterations. Those results are consistent with the results here except for the initial increase of errors that we observed. where the dark region has been overcompensated for.4. the cone artifact caused by the solid sphere inside the LROI is suppressed. . there seems to be a steady increase in reconstructible region during the first fifty iterations. the other artifact remains unaffected. After five iterations. 5. Fig. After five iterations. To examine the possibilities to reduce artifacts caused by objects located both inside and outside the LROI. In our case it was reasonable to use WFBP. 5.4 Experiments 85 80 75 70 65 RMSE (σe2 ). incorporates data outside the Tam-Danielsson window. Fig.52 β=0 β=0 β = 0. As the number of iterations further increase. 5. The image is distorted by artifacts caused by one solid sphere located in the LROI. this defect disappears. low contrast regions (HU) 25 20 15 10 5 0 0 63 RMSE (σe1 ). we have selected a cylindrical LROI just above the outer nine o’clock sphere in the clock phantom. A related study on iterative application of the PI-method [11.18: RMSE (σe1 and σe2 ) for head phantom reconstructions at a peripheral image not properly reconstructible with WFBP(z = 59mm).3. 5. [65]. 5.52 50 Iteration number 100 0 50 Iteration number 100 Figure 5. in contrast to the PI-method.21 c) and d) show WFBP and RIWFBP LROI reconstructions without the preprocessing of input data described in Fig.21 a) shows a full ROI WFBP reconstruction zoomed in over this LROI.4 Local region of interest The LROI reconstruction method was described in Sec. which shows axial images at z = 60mm. and one solid sphere located just below the LROI. 5. However. reconstructible with the WFBP method. This effect is even more visible in Fig. Obviously. the WFBP LROI reconstruction is identical to the WFBP full ROI reconstruction. an improved result is obtained. 110] was made by Magnusson et al. For the initial full reconstruction over ROI we are free to choose any reasonable non-exact algorithm. and turned into a slightly too bright region. One reason for this inconsistency could be that the WFBP method.20. the cone artifacts are suppressed and the windmill artifacts are slightly alleviated as shown in Fig. 5.

β = 0. Greyscale window ±50HU.52 g) RIWFBP. β = 0.52 e) RIWFBP. β = 0 h) RIWFBP. f50 . f5 .64 Regularized iterative weighted filtered backprojection a) Phantom b) WFBP reconstruction c) RIWFBP.19: Sagittal images of RIWFBP reconstructions of the head phantom. f5 . β = 0. β = 0 f) RIWFBP. f50 .52 Figure 5. f10 . . f10 . β = 0 d) RIWFBP.

For comparison. and the result look very similar to the full ROI result. Greyscale window ±50HU. De Man et al.5. and there is also a clear alleviation of windmill artifacts.21 e) and f) show WFBP and RIWFBP LROI reconstructions on LROI preprocessed input data. the pre-processing suppresses artifacts non-local artifacts caused by objects located outside the LROI. 5. current state of the .78◦ . Clearly. [18] state that for convergence. the other cone artifact has also been suppressed.5 Discussions and conclusions 65 a) WFBP reconstruction b) RIWFBP.5 Discussions and conclusions The main advantage of the proposed RIWFBP method is its high rate of convergence relative to other iterative methods.20: Axial images of RIWFBP reconstructions of the head phantom at z = 60mm. Fig. 5. f50 Figure 5. f5 c) RIWFBP. After five iterations. For full resolution medical CT input data at cone angles up to ±2. cone artifacts are perfectly suppressed within three iterations. f10 d) RIWFBP.

.. Greyscale window ±20HU. without preproc. f5 e) LROI. f5 c) LROI.. f0 f) LROI. with preproc. f5 Figure 5.66 Regularized iterative weighted filtered backprojection a) Full ROI..21: Full ROI and LROI reconstructions of the clock phantom. with preproc.. without preproc. f0 b) Full ROI. f0 d) LROI.

the weighting makes the corresponding optimization problem more difficult to solve within a short period of time. the dose utilization for an exact method is less than 50% for the Tam window. The Poisson model may be possible to adopt in RIWFBP. one might think that there would be no need for non-statistical iterative methods such as the RIWFBP method. [26]. Since exact methods directly produce results free from cone artifacts. masking. For some approximate methods (Flohr et al. since they are not based on iterative application of projection and backprojection operators.22). Software approaches include interpolation between conjugate rays during backprojection [38]. [88]. Zamyatin et al. since the principles of incorporation of redundant data are similar for the WFBP and the Adaptive Multiple Plane Reconstruction (AMPR) method (Flohr et al. However. Schechter et al.22. 5. [109] include weighting factors. and modified linear interpolation [119]. we believe that RIWFBP has at least three important advantages over currently presented exact methods.5 Discussions and conclusions 67 art methods for statistical reconstruction (iterative coordinate descent (ICD) [109] and ordered subsets convex (OSC) [52]) require a computation time equivalent to 60 RIWFBP iterations. and Zamyatin et al. On the other hand. 5. • Third. the signal to noise ratio can be improved. existing exact methods are limited to utilizing data from the Tam window [108]/PI detector [10] or the n-PI detectors [80] (see Fig. [26]). However. A number of other approaches exist for alleviating wind-mill artifacts in helical cone-beam CT. the RIWFBP method has the advantage of not only . but it is far from obvious how such weighting can be incorporated into the RIWFBP framework without increasing the number of iterations considerably. which take into account the Poisson distribution of measured input data. Recently. by Thibault et al. we believe that this is true also for the WFBP method. and Hilbert transformation characteristic for the proposed exact methods. [120]). Also. Hereby. [114] presented exact reconstruction methods based on work by Katsevich [48] and Zou and Pan[128] as potential alternatives to contemporary methods in commercial systems. [120] experimentally verified that this affects the signal to noise ratio negatively. the iterative nature of RIWFBP makes it possible to incorporate more advanced forward projection models in order to compensate for blurring effects caused by the finite size of focal spot and detector elements. Statistical reconstruction methods studied e. As shown in Fig. Compared to the RIWFBP method.g. RIWFBP does not suffer from the sensitivity in implementations of derivation. Wang et al. and the gantry rotation during acquisition (see the final paragraph). • First. for a pitch factor of 0. • Second.5. these methods are faster. non-linear filtering [35].75. it has been demonstrated that incorporation of data outside the Tam window can improve the signal to noise ratio.

68 Regularized iterative weighted filtered backprojection z y x Figure 5. but also suppressing cone artifacts. Dark gray: PI-detector. alleviating wind-mill artifacts. In this chapter. and other types of high-pass filters. the PI detector and the 3-PI detector for a pitch factor of 0. because of the full linearity of each and every part of the reconstruction system. there also exists a hardware technique already employed in clinical CT systems [27]. Hence. Therefore. and increasing the number of readouts per turn with a factor of two. we have limited our investigations to linear regularization because it simplified the convergence analysis. spatial resolution measures such as MTF and SSP can be used for analysis and characterization of the reconstruction method. An important issue in iterative reconstruction is linear versus non-linear regularization. Unfortunately.75. Furthermore. all of which could be used to further equalize the frequency response between WFBP and RIWFBP. and there are many other possibilities. a virtual detector with twice as many rows is created. making it more suitable for systems with higher cone angles. There is no strong theoretical motivation for this choice. software solutions for alleviating wind-mill artifacts are still an interesting topic of investigation. Another benefit from the linearity is the possibility to . the moving spot mechanism comes with increased complexity and costs in many parts of the acquisition system. Light gray: clinical CT detector.22: Illustration of the relation between a clinical CT detector. for instance more accurate Laplacian approximations. Our investigations employed a 3 × 3 × 3 approximation of a Laplacian for regularization. White: 3-PI detector. resulting in a considerable alleviation of wind-mill artifacts. It should be mentioned that besides the above mentioned software approaches for wind-mill artifact reduction. By rapidly moving the focal spot between two different z-positions. the aliasing during acquisition is reduced.

4. As shown in Section 5. the update step in RIWFBP is completely simultaneous. the resulting method is highly parallelizable. Thibault et al. if any contemporary reconstruction methods include compensation of these effects. In iterative reconstruction. similar variations are expected for noise and SSPs. and can be efficiently implemented in modern hardware. non-linear regularization has shown to be useful for preserving high-contrast resolution while reducing noise at low contrasts (see e. These variations follow the variations of the WFBP MTF. Few. and will therefore be the topic of Chapter 8. g. there are spatial variations in the MTF of the RIWFBP method. and the problem of finding the best projection/voxel access scheme is non-existent. the finite size of the detector elements. within each iteration. [109]). it is far more easy to incorporate accurate acquisition models. Therefore. Although not investigated here. and the gantry rotation during the measurement. In other words. Furthermore.5. measured data are blurred by the finite size of the focus spot. In contrast to column-action methods such as the ICD [109] method or rowaction methods such as the algebraic reconstruction technique (ART) [31].2. there are no dependencies between updates of different voxels or rays. the problem of finding efficient relaxation factors is less critical. During the input data acquisition.5 Discussions and conclusions 69 use resolution modifying filters frequently used in clinical CT systems. which results in reconstructed images with spatial resolutions below the ones to be expected from actual numbers of views and detector element densities. Nonetheless. Improving the acquisition model implemented in the forward operator P is the topic of Chapter 9. .

70 Regularized iterative weighted filtered backprojection .

we summarize the conclusions and discuss possible future investigations. 6. Chapter 6]. . SART. we first describe the OS technique and then move on to how it is applied in the OS-IWFBP method.8◦ ).Chapter 6 Ordered Subsets IWFBP (OS-IWFBP) 6.1 Introduction This chapter is an adapted and reduced version of [102. presenting the OS-IWFBP method. many recent investigations indicate a need for iterative methods with a higher rate of convergence than the RIWFBP from the previous chapter. our investigations have focused on other topics. g Mathematical descriptions and proofs of convergence for sequential methods such as ART. In those cases. Later. Since the original publication. which are unlikely to be used for medical helical cone-beam CT. the Ordered Subsets (OS) technique could be an interesting candidate for improving the rate of convergence. This is followed by an experiment showing how OS-IWFBP efficiently suppresses cone artifacts for high cone angles. mainly because the OS-IWFBP method seemed to be useful only for extremely high cone angles (κmax > 4. In the following sections. They showed that this technique could improve the rate of convergence for the emission tomography MLEM method [90] with at least a factor 16 without any appreciable loss of image quality.2 The Ordered Subsets (OS) technique The OS technique for emission tomography was introduced in 1994 by Hudson and Larkin [39]. the ordered subsets technique was applied to algorithms for transmission tomography by Kamphuis and Beekman [47] and Erdo˘an and Fessler [24]. and OS methods were given by Jiang and Wang in [40]. However. Finally. which is a non-regularized and accelerated version of the RIWFBP method. Here.

1).3 respectively.1) converges towards a minimum norm minimizer of Pf − pin 2 plus a projection of the initial image f0 on the null space N (P) = R(PT )⊥ . followed by a presentation of the OS-IFBP method. A simultaneous method would use all projection data pin in each update. 6. Each arrow represents a parallel projection angle θ.2) then the method defined by (6. the input data pin are partitioned into equidistantly sampled subsets pin.}) should be avoided. the update step becomes fk+1 = fk + αk PTk (pin. . With an index sequence {ik }∞ specifying the order in which the subsets shall k=1 be applied. .72 Ordered Subsets IWFBP (OS-IWFBP)                             Figure 6. such as random ordering. the full update would be split into three partial updates working on pin.. there are very small differences between common ordering schemes. we give a brief presentation of the OS technique for least squares minimization (see Section 3.1)..1 ..1) For the case of disjoint subsets. a projection matrix Pl is defined.2 . and the ordering scheme by Beekman and Kamphuis [2] which we will use in the following sections. This matrix only generates projection data for the subset l. In an OS method. and k=0 αk = +∞. .l . n ∈ N. i (6. 2.ik − Pik fk ). and pin. Previous investigations have shown that while obviously inefficient orderings such as linearly increasing θ-offset ({ik } = {1. Jiang et al. (6. i. l = 1.. the golden ratio based ordering [49]. 3.e. and the step lengths αk satisfy the conditions ∞ k→∞ lim αk = 0.1: Simple illustration of OS in the case of 12 projection angles and three subsets. pin. [40] showed that if the index sequence is periodic. ik+nL = ik . For each subset.e. In the illustrated OS case.. i. Although the above result holds independently of subset ordering. the ordering affects the rate of convergence. Pl consists only of the rows of P corresponding the subset l. L with respect to the rotation angle θ (see Fig.

.3) Assuming that the ordering {ik } is periodic with period L. (6. Scanning geometry parameters for all experiments in this chapter are listed in Table 6. Given a number of subset indices i1 . and step length αk = α.3 OS-IWFBP 73 6.. the . certain noise-like artifacts appear in the OSIWFBP results (see Fig. the next index in+1 is chosen as close as possible to the center of the largest gap of unused subsets.4) where C = m exp − m2 so that the DC-component of the filter equals one. Visual inspection reveals that for high values of α. These data are then divided into L subsets preb.e. i2 . Sometimes several subsets satisfy the last condition (see for instance Fig... 6. σ We use the ordering scheme suggested by Beekman and Kamphuis [2].6 c)). For each subset.. we have performed two-dimensional experiments with the Forbild thorax phantom [97] shown in Fig. as we will see in the next section. m = − M2 . The kernel used for low-pass filtering is a truncated and sampled Gaussian given by Kσ [m] = 1 C exp − m2 σ 2 2 0 −1 −1 . .L .2.ik − Pik fk ). The underlying assumption is that unwanted artifacts generated by the OS technique will prevail in the same way in the three-dimensional case. as close to LΔθ /2 as possible.15 for ten subsets). M2 . in . . If several gaps of the same size exist. divergent high frequency structures develop during the iterations.. 6. . otherwise (6. preb. i.3. and that all subsets are applied during one period.2 e)). row-wise low-pass filtering of projection data differences (preb − Pik fk ) is incorporated in the reconstructions Ql . To allow for higher step lengths. 6. Ten full iterations per configuration have been performed. we define one full iteration as the collection of L consecutive sub-iterations. The iteration of OS-IWFBP is given by fk+1 = fk + αk Qik (preb. However. is chosen.3 OS-IWFBP Similarly to the RIWFBP.. the center of the gap located as far away as possible.4 Experiments To examine how reconstruction errors depend on the number of subsets L. unless very low step lengths αk are used (αk = 0. 6. 6. Then the choice between the remaining centrally located subsets is random.1.6..1 . the first step in the OS-IWFBP method is rebinning of the input data pin to semi-parallel data preb . a specific projector/reconstructor pair Pl /Ql is created. As a numerical measure of these artifacts. High values of α (close to one) imply efficient suppression of cone artifacts. This ordering scheme is illustrated in Fig...

.3} ⇒ i2 = 2 c) i2 = 2 ⇒ largest gap = {5.3: Axial image of the Forbild thorax phantom [97] at z = 0mm. Figure 6..2: Illustration of the ordering scheme suggested by Beekman and Kamphuis [2]. the largest gaps of unused subsets are identified. The new index in+1 is chosen as close as possible to the gap center located as far away from the subset in as possible. Given a number of subset indices i1 ...6} ⇒ i1 = 4 b) i1 = 4 ⇒ largest gap = {1. {3}. {5} ⇒ i4 = 3 e) i4 = 3 ⇒ equally sized gaps = {1}.74 Ordered Subsets IWFBP (OS-IWFBP)        a) i0 = 0 ⇒ largest gap = {1... in ..6} ⇒ i3 = 6 d) i3 = 6 ⇒ equally sized gaps = {1}... i2 . .. {5} ⇒ i5 = 1 or i5 = 5 Figure 6.

192mm 0◦ /4.4). Error measurements σe with respect to the number of iterations are shown in Fig. Therefore. Source-Isocenter distance Number of channels (Quarter offset) Number of rows Number of projections per turn Detector height Table-feed Maximal cone angle error (repetition of (4. 96mm. For the different subset configurations included in this experiment. 2. To examine the effect of the lowpass filter in (6. These images show that for α = 0. The σ-values determining the amount of low-pass filtering were set to 1.4). and 4.5 mean in terms of visual artifacts. Therefore.4) with σ = 1. 6.4 Experiments Table 6. the OS-IWFBP produces similar images to those of the RIWFBP without regularization.8◦ .25. For five subsets with α = 0. strong ringing artifacts appear in the result.5) has been calculated over low contrast regions. 192mm 0mm. When σ is further increased. experiments have been performed with five and ten subsets. the amounts of artifacts at low contrast regions are reduced as σ is increased.15. to see what the numbers presented in Fig. However.5)) σe = 1 |Ω| ((frec )i − (fphan )i )2 i∈Ω 75 RS Nch Nrows Nproj h P κmax 570mm 336 1/64/128 580 1.4 and 6.15 to 0. the previously divergent sequences are turned into more stable sequences showing no signs of divergence during the first ten iterations.8◦ and one 128-row system with . the black solid line with ring markers show the corresponding error for RIWFBP with β = 0. To study cone artifact suppression.50 and ten subsets with α = 0. when α is increased from 0. The RMSE gives no information about the structure of the errors. two systems were considered: one 64-row system with a maximum cone angle κmax = 4.6. In all four cases. axial images of reconstructions with ten subsets and two full iterations are shown in Fig.25. 6.6. For comparison. the highest allowable α varies strongly with the number of subsets.5mm. for large σ values we expect the σe values to slowly increase to levels similar to those of the RIWFBP with β = 0.5. Clearly. 9. the amount of artifacts at low contrast regions drops below those introduced by RIWFBP with β = 0. This is mainly because increasing σ also increases the required number of iterations for high frequencies to be reconstructed.1: Scanning parameters for experiments on OS-IWFBP. 6. α = 1/L seem to result in approximately the same error curve as the RIWFBP method with β = 0.4 shows σe measurements for different number of subsets and step lengths without the lowpass filter in (6. these artifacts are clearly suppressed. 96mm.6◦ (6. 6. Fig. By using the lowpass filter from (6.

α=1. α=0. α=0. α=0.5 3 2.25 OS−IWFBP.00 OS−IWFBP.2 4 RMSE (HU) 3.00 OS−IWFBP.5 2 0 b) 5 subsets RIWFBP.07 OS−IWFBP.5 3 2.10 4 RMSE (HU) 3.15 OS−IWFBP. α=0.4: RMSE/σe measurements on low contrast for RIWFBP with β = 0.50 OS−IWFBP. α=0.35 OS−IWFBP.5 4 RMSE (HU) 3. α=0. α=0. α=0. α=0.76 Ordered Subsets IWFBP (OS-IWFBP) RIWFBP. α=0. and OS-IWFBP with different number of subsets and step lengths.02 2 4 6 8 Number of full iterations 10 c) 10 subsets d) 29 subsets Figure 6.10 OS−IWFBP. α=1.0 OS−IWFBP.0 OS−IWFBP.00 OS−IWFBP. α=0.5 3 2.5 2 0 RIWFBP. α=0.5 2 0 2 4 6 8 Number of full iterations 10 OS−IWFBP.05 OS−IWFBP. .5 2 0 2 4 6 8 Number of full iterations 10 OS−IWFBP.10 2 4 6 8 Number of full iterations 10 a) 2 subsets RIWFBP.7 OS−IWFBP. α=1.20 OS−IWFBP. α=1. α=1. α=0. α=0.5 3 2. α=0.05 4 RMSE (HU) 3.

α=0. α=0. σ=1 OS−IWFBP. σ=4 4 RMSE (HU) 3. No filter OS−IWFBP.25.5 2 b) 5 subsets. σ=4 0 2 4 6 8 Number of full iterations 10 c) 10 subsets.6.00 OS−IWFBP. σ=2 OS−IWFBP.5 3 2.5 3 2.35 RIWFBP. α=0.5. α=0. α=0. σ=1 OS−IWFBP.15.35. σ=1 OS−IWFBP. α=0. σ=2 OS−IWFBP. No filter OS−IWFBP.15.25 Figure 6. α = 0. No filter OS−IWFBP.25. α=0. α=0.5 3 2.4 Experiments 77 RIWFBP. σ=1 OS−IWFBP.5 2 0 RIWFBP.5 2 0 2 4 6 8 Number of full iterations 10 OS−IWFBP. α=1.00 OS−IWFBP. α=0.15.15. α = 0. α = 0.25. α=0. α=1.5. α=0.5. α=0.50 RIWFBP.15 d) 10 subsets. α=1. No filter OS−IWFBP.5 2 0 2 4 6 8 Number of full iterations 10 OS−IWFBP. α=0.35. σ=2 4 RMSE (HU) 3.35. σ=4 4 RMSE (HU) 3.00 OS−IWFBP. σ=2 4 RMSE (HU) 3. α=0.5 3 2.35.5: RMSE/σe measurements on low contrast regions for different degrees of low-pass filtering. .00 OS−IWFBP. α=1.5. α = 0.25. α=0. σ=4 2 4 6 8 Number of full iterations 10 a) 5 subsets. α=0.

α = 0.15.25. and how they can be suppressed by the low-pass filter in (6.25.4).6: Blown up two-dimensional reconstructions of the Forbild Thorax phantom [97]. β = 0 b) OS-IWFBP. showing high-frequency “ringing” artifacts. Greyscale window ±50HU. No LP c) OS-IWFBP.78 Ordered Subsets IWFBP (OS-IWFBP) a) RIWFBP. α = 0. . The number of subsets is ten.0. No LP d) OS-IWFBP. σ = 1 Figure 6. α = 1. α = 0.

for κmax = 9. this increase in computational cost per iteration needs to be taken into account when examining improvements in rate of convergence. for a low number of subsets L. it is difficult to motivate the use of OS for improved cone artifacts suppression for low cone angles. the maximum allowable step length seems to be inversely proportional to the number of subsets. if the computational cost of R is one tenth of the cost of Pl and Ql . However. Our results show that by applying a Gaussian lowpass filter with σ = 1 in the update step. For example. The OSIWFBP result is only marginally better.5 Discussions and conclusions 79 a maximum cone angle κmax = 9. However.6◦ shown in Fig. Already after one iteration. while the image produced by OS-IWFBP is close to perfect in this respect.8. 6. After two iterations. for the higher cone angle ±9. or in other ways improve the image quality.5 Discussions and conclusions The main motivation for combining the OS technique with the RIWFBP algorithm is to reduce the required number of iterations to suppress cone artifacts. resulting in the update step fk+1 = fk + αk Qik (pin.6) This is motivated by recognizing that Ql ≈ Q. as the number of subsets grows.8◦ . As shown in the previous section. OS-IWFBP is not convergent. and from 0. One way to allow for higher step lengths and thereby improve the efficiency is to lowpass filter projection data in the update step according to (6. the OS technique does not seem to offer much improvement to the result obtained with RIWFBP. the step length α should be close to one.6. 6. there are clearly visible cone artifacts remaining in the RIWFBP result. . we note that it can be added in the same way as for the RIWFBP.15 to 0.50 for five subsets. the differences in efficiency are less pronounced. In this case.25 for ten subsets.ik − Pik fk − βRfk ). the approximation becomes more inaccurate.6) is approximately equal to the RIWFBP update step. the OS-IWFBP efficiently improves the reconstruction results during the first iterations.4). However. the update step (6. Obviously. 6. Thus. Instead. Therefore. decent suppression of cone artifacts is obtained without OS. the difference between the methods is more pronounced. Although regularization has not yet been investigated for the OS-IWFBP. To obtain maximal suppression of cone artifacts. the cost of a full iteration increases with 50% for ten subsets and 100% for twenty subsets. However. since there is no mechanism to control the step lengths of the method.7 shows axial reconstructions of the clock phantom for κmax = 4. Indeed.6◦ the OS-IWFBP clearly performs better than the RIWFBP method in terms of cone artifact reduction. (6. for cone angles smaller 4.6◦ .35 to 0.8◦ . and multiplication with R becomes more expensive relative multiplication with Pl and Ql . Fig. However. the maximum allowable α can be increased from 0.

f0 (c) RIWFBP. f1 (d) OS-IWFBP. β = 0. For the OS-IWFBP reconstructions. f0 (b) OS-IWFBP. and for e) and f).7: Clock phantom reconstructions with RIWFBP (β = 0) and OSIWFBP with 10 subsets.80 Ordered Subsets IWFBP (OS-IWFBP) (a) RIWFBP. β = 0. β = 0. Note that the computational cost is approximately the same for c) and d). Greyscale window ±50 HU. f10 (e) RIWFBP.25 and σ = 1. .8◦ . f2 (f) OS-IWFBP. α = 0. f20 Figure 6. Cone angle κ ≈ ±4.

f2 (f) OS-IWFBP. β = 0.6◦ . f10 (e) RIWFBP. f1 (d) OS-IWFBP. . β = 0. f20 Figure 6.6.8: Clock phantom reconstructions with RIWFBP (β = 0) and OSIWFBP with 10 subsets. Greyscale window ±50 HU. and for e) and f). Cone angle κ ≈ ±9. β = 0. Note that the computational cost is approximately the same for c) and d). f0 (b) OS-IWFBP.5 Discussions and conclusions 81 (a) RIWFBP. f0 (c) RIWFBP.

it is shown that these techniques can improve images in terms of spatial resolution versus noise. However. this comes to the cost of an increased number of required iterations.82 Ordered Subsets IWFBP (OS-IWFBP) it either diverges. . In Chapter 8 on nonlinear regularization and Chapter 9 on improved data acquisition modeling.e. or enters a limit cycle.. i. an interesting topic for future research is to examine if OS can be used to improve the rate of convergence in these cases. and then either to switch to another method that guarantees convergence [98]. or to successively reduce the number of subsets [2]. One way to avoid these phenomenons is to use the OS-IWFBP for a couple of initial iterations. an infinite cycle of recurring reconstruction results. Therefore.

the method is less satisfying. Q.2 are unneccesarily inefficient and complicated to implement.Chapter 7 Relation between RIWFBP and least squares methods 7. Thus. If there is a fixed point of the RIWFBP iteration. and QT in each iteration. the RIWFBP method is a practical method for improving reconstructed images within very few iterations. already f0 = Qpreb corresponds to a relatively good reconstruction result. 2 2 (QP + βR)T ((QP + βR)f − Qpreb ) T T T T 2 2 T (7. any gradient based method for minimizing z(f ) would require calculation of matrix-vector products with P.e. the gradient of z(f ) equals ∇z(f ) = = (7. standard tools from optimization theory for minimizing z(f ) in 7. we suggest a new method called WLS-IFBP. for which such comparison is possible. it can be expressed as (repetition of (5. . Therefore. it is difficult to relate and compare RIWFBP to other iterative methods. i. Still. In particular.1 Motivation The RIWFBP method presented in Chapter 5 is intuitively appealing: since Q is an approximate inverse of P. from a theoretical point of view. noise.2) Assuming that R is symmetric.1) (7.3) (P Q QP + βP Q R + βRQP + β R )f − (QP + βR) Qpreb . and spatial resolution.5)) f∞ = (QP + βR)−1 Qpreb which trivially minimizes the quadratic cost function z(f ) = 1 (QP + βR)f − Qpreb 2 . To this end.. This new method is compared to the RIWFBP method with respect to artifact reduction. PT .

6) is satisfied if and only if the matrix SQP can be written as VBP. and to use tools from optimization theory for better convergence analysis and improvements.84 Relation between RIWFBP and least squares methods In contrast to RIWFBP. we would be able to more easily compare it to competing methods. Therefore. Comparing (7. e. and measures the error in the projection domain. a factorization BP = PT WP exist. where ˆ ˆ BP is symmetric and non-negative. . we have examined the possibility to express the RIWFBP update step fk+1 = fk − αS(Q(Pfk − preb ) + βRfk ). 7. (7. the WFBP method employs a backprojection angle dependent normalization (see Chapter 2). Therefore. a rampfilter. and use of more advanced gradient based methods such as the conjugate gradient method.2 Weighted least squares IFBP (WLS-IFBP) As shown in the previous section. as a gradient descent step modified so that fk+1 = fk − αV∇z(fk ) (7.5) (7.5). which is similar to the RIWFBP method but has an update step that satisfies (7.7) However. and z(f ) is some quadratic function to be minimized. To find a quadratic cost function that is minimized by the RIWFBP method. most iterative methods are formulated as optimization problems in which the error is measured in the projection domain rather than the reconstruction domain. (7. where W is a symmetric non-negative weighting matrix. it is highly unlikely that the RIWFBP method could be expressed as in 7.5). it is difficult to analyze the RIWFBP method in terms of a least squares problem.5 for any quadratic cost function.g. the corresponding cost function would be z(f ) = 1 ˆ Pf − preb 2 2 W + Regularization term. which makes it very difficult to express SQP as described above. Then.4) where V is symmetric and positive definite. we see that the difference image in RIWFBP must satisfy V∇z(f ) = S(Q(Pf − preb ) + βRf ). If we could reformulate the RIWFBP method in similar terms. In such case. we present and evaluate the Weighted Least Squares IFBP (WLS-IFBP) method. and does not involve an explicit reconstruction operator Q. To obtain a method that is easier to analyze. In the following sections.6) Any gradient of a quadratic cost function is symmetric and non-negative. (7.4) and (7. Examples of such improvements include automatic determination of step lengths for guaranteed convergence.

where H corresponds to a rampfilter. This is done to suppress artifacts associated with sharp transitions between voxels receiving contributions and voxels not receiving contributions.12) . where ·. · 2 is defined as W·. The downweighting function is the same as the one used for the WFBP method [100]. and WQ corresponds to a downweighting of the most peripheral detector rows. In (7. 2 (7.7. 2. we obtain the update step ˜ ˜ fk+1 = fk − αk V(Q(Pfk − preb ) + β Rfk ). · denotes the Euclidean scalar W product. As will be demonstrated in the next section.9) cos 2 1−Q .7. and the regularization operator is given by ⎞ ⎛ N N N ∂ ⎝ vi (fi − fj )2 ⎠ ˜ el dij Rf = ∂fl i=1 2 j=1 2 l=1 ⎛ ⎞ N N N vi vl ⎝ dil (fi − fl )⎠ el = dlj (fl − fj ) − 2 j=1 2 i=1 l=1 (7.2 Weighted least squares IFBP (WLS-IFBP) 85 we consider the problem of finding a vector/reconstructed image f minimizing the function z(f ) = 1 Pf − preb 2 N 2 W +β i=1 vi 2 N dij j=1 (fi − fj )2 . ⎪ ⎩ 0. change of summation variables} N N dij i=1 j=1 (vi + vj ) (fi − fj )ei . The coefficients vi are selected as the column sums of WQ P in order to approximately preserve the local spatial resolution properties of the RIWFBP method.10).11) = = {dij = dji . By differentiation of z(f ) and multiplication with the positive definite matrix √ √ V = diag(1/ vi ) × S × diag(1/ vi ).10) (7.8) which is a special case of the generally stated quadratic cost function proposed by Delaney and Bresler [22]. 2 (7. and is given by ⎧ |q| < Q ⎪ 1. 1] is the detector row coordinate. normalized such that q = ±1 means the bottom and top boundaries of the detector respectively (see Fig. a reasonable number for the parameter Q is approximately 0. Q ≤ |q| < 1 . ⎨ 2 π |q|−Q WQ (q) = (7. ˜ where Q = PT W. |q| ≥ 1 where q ∈ [−1. · . The matrix W equals WQ H.8).

In this case of a quadratic cost function. i. variations between neighboring ˜ values vi are small in most cases. . the cost function can also be written as z(f ) = 1 ˆ W(Pf − preb ) 2 2 2 + β ˆ Rf 2 2 2 = 1 2 ˆ ˆ WP √ ˆ f − Wpreb 0 βR =:A =:b 2 . To stay consistent with the method and experiments presented in the previous chapter. we make one final remark regarding con˜ vergence and choice of step lengths αk .. the WLS-IFBP method is initialized with WFBP. it can be factorized as R T ˆ ˆ way. f0 = Qpreb .4). Since the symmetric matrix R is weakly diagonally ij ˜ = RT R. Comparing (7. (7. this choice yields a globally convergent algorithm for minimizing z(f ). As shown in several textbooks. However. Thus. ⎠ ˜ ˜ dN j − d N N (7.11) and (7. W can be factorized as W = W W. . pT AT Apk k (7. this automatic determination of αk can be made computationally inexpensive. we do not use this automatic determination of α in the experiments that follow. Nocedal and Wright [71].14) 2 The step length αk for calculating fk+1 can now be chosen as the global minimizer of z(α) = z(fk + αpk ) where pk = −V∇z(fk ). this operator produces results corrupted by severe low-frequency artifacts when applied as a direct reconstruction operator. ⎟ . . The matrix R can explicitly be written as ⎛ ⎜ ⎜ ⎜ ˜ R=⎜ ⎜ ⎜ ⎝ N j=1 d˜ − d˜ 1j 11 −d˜ 12 N j=1 ··· .g. In the same ˆ ˆ dominant. ˜ −dN 1 d˜ − d˜ 2j 22 . and thus non-negative. N j=1 ˜ −dN 2 ··· ⎟ ⎟ ˜ −d2N ⎟ ⎟ ⎟ . Therefore. it requires additional operations equivalent to one iteration to be performed prior to starting the first iteration. Therefore. Since the normalization in VQ allow certain angles to contribute more than others (depending on the local amount of data redundancies). Before continuing with experiments. in order to reduce the number of required iterations.15) By reusing results. Instead we use a constant step length αk = α as determined in the previous chapter.13) ˜ −d1N ⎞ ˜ where d˜ = dij (vi + vj )/2. this step length can be calculated analytically as αk = −pT ∇z(fk ) k . we see that the WLS-IFBP method can be ob˜ ˜ tained from the RIWFBP method by replacing SQ and SR with VQ and VR re˜ spectively.86 Relation between RIWFBP and least squares methods Except for the outer parts of the field of view. V ≈ S × diag(1/vi ) and VR is well approximated by SR. e. . −d˜ 21 .e.

101]. employing bilinear interpolation adapted to the detector sampling. 7. and b) the “Standard” backprojector BS . β = 0.2 b) shows a non-iterative WFBP reconstruction of the clock phantom. The regularization parameters previously determined to give similar resolution properties for the RIWFBP and WFBP method were used . 7. 7. Fig. Fig. we used the “standard backprojector” BS . and (ii) windmill artifacts showing up as alternating bright and dark regions near sharp edges orthogonal to the z-direction of the reconstruction volume. two types of artifacts are present: (i) lowfrequency cone artifacts caused by the non-exactness of the reconstruction method. The Joseph backprojector PT does not satisfy this criterion.7.1: WLS-IFBP reconstructions of the clock phantom using a) the Joseph backprojector PT . In any forward projection or backprojection operation.0. or equivalently with symmetric forward projector/backprojector pairs.3 7.1 Experiments Artifact reduction We begin by comparing artifact suppressing abilities of the RIWFBP and the WLS-IFBP using noise-free data. Note that results obtained with the Joseph backprojector.52. 7. The topic of improving the projection operator P is revisited in Chapter 9.7. f5 Backprojector: BS Figure 7.2 c)-f) show results after one and two iterations with the RIWFBP and WLS-IFBP methods. βxy = 1. 7. Greyscale window ±20HU. βz = 1.3.5.3 Experiments 87 a) WLS-IFBP. can be improved for instance by modeling one detector measurement as a mean value of several line integrals instead of just one. the interpolation function must be adapted to the input sampling grid in order to avoid DC-aliasing [12. To avoid these artifacts. f5 Backprojector: PT b) WLS-IFBP. As discussed in the previous chapter. thus producing severe artifacts when used as backprojector as shown in Fig.1 a). Reconstruction parameters: Q = 0. Fig.1 b) shows the corresponding result obtained using BS .

7.6.5). For the iterative reconstructions. βz = 1.4 shows WFBP. As demonstrated in the previous chapter.0 from 0. Only a very small reduction of the error over the whole image is observed.0 with non-iterative WFBP results in strong non-local artifacts. MTF measurements for the RIWFBP and the WLS-IFBP method are shown in Fig. . 7. Q = 0. the relative difference is again very small.2 Spatial resolution and noise Spatial resolution has been examined by measuring MTFs and SSPs with the method described in Section 4.0. the distance to isocenter does not seem to affect the SSPs much. or increase the noise level with 5% (Q = 0. 7. when the two methods have nearly converged. 7. In relation to WFBP. and depending on the choice of Q. Obviously. In this particular experiment.7. In contrast to the xy-plane resolution. using Q = 1.0). and at a distance of 150mm away from the isocenter. an error reduction from 5HU to 2HU is observed for all six reconstructions. also the SSPs for RIWFBP and WLS-IFBP are very similar. the errors seem to get lower for Q = 1.0.0 than for Q = 0.3).88 Relation between RIWFBP and least squares methods (β = 0.7.3.3.0. RIWFBP.7 and Q = 1. the iterative methods may decrease the noise level with 5% (Q = 1.3. However. preserve the noise level (Q = 0. Again.7). 7.3. Again. the RIWFBP method suppresses cone artifacts slightly faster than the WLS-IFBP method. After five iterations. Fig. 7.8 shows noise measurements on clock phantom reconstructions for RIWFBP and WLS-IFBP at approximately equal spatial resolution. using edge phantoms of radius 20mm placed in the isocenter. Fig. The MTFs for RIWFBP and WLS-IFBP are nearly identical irrespective of distance to isocenter. The errors σe1 (whole image) and σe2 (without edges) are shown in Fig. there are only small differences remaining in the close vicinity of the outer solid spheres as shown in Fig. Also for Q = 1.52. containing only low contrast structures (Ω2 ). some new artifacts are introduced in the vicinity of the outer spheres as the downweighting parameter Q is increased to 1. Over regions in the image. βxy = 1. there is a large difference in MTF depending on the distance to the isocenter. and error measurements of head phantom reconstructions show the opposite relationship between the errors. 7. the artifacts are slightly more pronounced for the WLS-IFBP method than for the RIWFBP method. For both methods. As expected. there are only small differences between the RIWFBP and WLS-IFBP reconstructions.5.7 was used for initialization. the RIWFBP and WLS-IFBP results look very similar. However. and WLS-IFBP reconstructions for Q = 0. the difference between the two methods is very small. The choice of Q seem to be more important. As shown in Fig.

Greyscale window ±20HU. βz = 1.0. f1 e) RIWFBP. βxy = 1.5.52. f2 Figure 7. Reconstruction parameters: Q = 0.3 Experiments 89 a) Sampled phantom b) WFBP reconstruction c) RIWFBP.7.7. . f2 f) WLS-IFBP.2: RIWFBP and WLS-IFBP reconstructions of the clock phantom. f1 d) WLS-IFBP. β = 0.

the similarity in formulation between WLS-IFBP and most other iterative reconstruction methods potentially makes it easier to relate to and use ideas and techniques from other methods.2. in its current form. The linear least squares formulation of the reconstruction problem makes it easy to analytically determine the optimal choice of step length α. i. f5 d) WLS-IFBP. Greyscale window ±20HU.7. 7. βz = 1. a small modification of the method results in the WLS-IFBP method.52. However. f5 c) RIWFBP. as shown in Section 7.90 Relation between RIWFBP and least squares methods a) RIWFBP. From a practical point of view.0. Furthermore. f5 b) WLS-IFBP. βxy = 1.4 Discussions and conclusions The RIWFBP method cannot easily be analyzed or understood with the tools from optimization theory.e. Reconstruction parameters: Q = 0. β = 0. which converges to the optimum of a weighted least squares problem provided that reasonable step lengths αk are used.3: RIWFBP and WLS-IFBP reconstructions of the clock phantom after five iterations. the step length that minimizes the cost function along the search direction.5. f5 Figure 7. the WLS-IFBP does not .

Regularization parameters: β = 0. Q = 0.0. f5 . βxy = 1.0 Figure 7. RIWFBP.7.0 e) WLS-IFBP. Q = 0. βz = 1.7 and Q = 1. Q = 1.7 f) WLS-IFBP.52. Q = 1. All iterative reconstructions have been initialized with a WFBP reconstruction using Q = 0. f5 . Greyscale window ±20HU. f5 .0.4 Discussions and conclusions 91 a) WFBP recon. Q = 1. Q = 0.7 d) RIWFBP. and WLS-IFBP reconstructions for Q = 0.. f5 .7 b) WFBP recon. .7.4: Comparison between WFBP.0 c) RIWFBP.5..

. Q=0.Q=1.1 0 0 1 2 3 4 5 6 Frequency (lp/cm) 7 8 9 10 RIWFBP. 7.5 0. r = 150mm Figure 7.3 WLS-IFBP.9 0. without edges (HU) 4 3 2 1 0 0 RIWFBP.8 0 5 Iteration number RIWFBP.2 50. whole image (HU) 10 5 Iteration number 10 Figure 7.5: σe measurements on the clock phantom reconstructions in Fig.0 RMSE.2 0.8 0.3.9 49.7 5 RMSE. The RIWFBP and WLS-IFBP curves overlap both in the isocenter and off-center cases. 1 0.Q=0. f 5 . Q=1.2 and 7. f 5 .7 0.7 WLS-IFBP. Q=0.3 50.1 50 49.3 0. r = 150mm WLS-IFBP. r = 0mm RIWFBP.4 0. f 5 .6 MTF 0. f 5 .Q=0.92 Relation between RIWFBP and least squares methods RIWFBP.6: MTF measurements of the RIWFBP and the WLS-IFBP method in the isocenter and 150mm off-center. r = 0mm WLS-IFBP.0 WLS-IFBP.3 50.

8 0.5 15 14.7 RIWFBP.3 WLS-IFBP. Q = 0.2 0.9 0.5 17 16.8: σn measurements made on clock phantom reconstructions. r = 150mm WLS-IFBP. f 5 . f 5 . . r = 0mm WLS-IFBP.7 Normalized SSP 0.7: Normalized SSP measurements of the RIWFBP and the WLS-IFBP method in the isocenter and 150mm off-center.5 0 1 2 3 4 5 6 Iteration number 7 8 9 10 RIWFBP.4 Discussions and conclusions 93 1 0.6 0.0 WLS-IFBP.7 WLS-IFBP.5 0. r = 0mm RIWFBP. 18 17. Q = 0. (HU) 16 15.3 0.5 RIWFBP.4 0. Q = 1.3 RIWFBP. f 5 . r = 150mm 1 1. Q = 0.5 Figure 7. Q = 0.5 Distance (mm) 2 2.1 0 0 0. f 5 .5 σn .7. Q = 1.5 14 13.0 Figure 7.

given a target PSF.g. to obtain uniform resolution). this difference in artifact reduction is smaller than expected. In contrast.94 Relation between RIWFBP and least squares methods perform as good as the RIWFBP in terms of artifact reduction.e. modeling one ray with the triangle function implied from the linear interpolation. has a regularizing effect on the result. considering the difference in how normalization is done in the two methods. Joseph with only one line integral per ray. Chapter 3) in this case is one single sharp line per ray. i. and is primarily motivated by preservation of spatial resolution properties of RIWFBP and WFBP. As shown in Chapter 9. For the projection operator used in this chapter. using PT as a backprojector results in an increase of certain artifacts. The choice of regularization parameters dij and vi presented in this chapter is ad hoc. This is because the irradiation function (cf. determine the optimal regularization parameters. an interesting future topic of research would be to investigate how such parameter choice methods can be adapted to the methods presented here. . the “standard” backprojector BS . thus avoiding increased artifact levels. Recent work by Shi and Fessler [91] in quadratic regularization for statistically weighted reconstruction aim at. Since this could provide means of controlling the spatial resolution (e. However. the Joseph backprojector produces better results when more line integrals are used for modeling one ray.

We conclude by discussing the results and presenting possible future research directions. (8. Nevertheless. 121. The next section presents non-linearly regularized versions of the RIWFBP and WLS-IFBP methods. employed to define the non-linearity of the regularization operator. This property simplifies the analysis and makes the results easy to predict: ignoring non-linearities in the acquisition. we consider the following cost function N vi N 1 zV (f ) = Pf − preb 2 + β dij V (fi − fj ). introduction of certain types of nonlinearities in the reconstruction can improve image quality.1) W i=1 2 j=1 2 . i. One important example is statistical reconstruction. This is motivated by the fact that the local signal to noise ratio is higher in high contrast structures than in low contrast regions.2 Description of the reconstruction method To incorporate nonlinear regularization in the RIWFBP and WLS-IFBP methods. In this chapter. 109. and so called potential functions.1 Introduction The methods studied in the previous sections are linear. any reconstructed image f depends linearly on input data p. Having obtained an understanding of the behavior of these methods for some parameter choices. where the signal to noise ratio is improved by taking the Poisson noise characteristics of input data into account [21. 126.. we study reconstructed images and noise measurements. we examine the possibility to use nonlinear regularization in RIWFBP and WLS-IFBP. 8. This is followed by a study of spatial resolution for different contrast levels. the reconstruction of any object is then independent of local contrast level and existence of other objects in the vicinity.Chapter 8 Nonlinear regularization 8.e. Many statistical reconstruction methods incorporate nonlinear regularization that reduces noise and spatial resolution for low contrast while preserving high contrast structures. 98].

96 Nonlinear regularization which is a generalization of (7. and for q = 1. this function becomes a quadratic function.1. Obviously.1. 1 + |f /c|p−q (8. the q-GGMRF functions are shown for different values of q. The q-GGMRF influence function is given by V (f ) = |f |p−1 1 + |f /c|p−q p− p−q cp−q |f |p−q 1 + |f /c|p−q sgn(f ). The function V (f ) is called potential function. the value of c determines where the intersection between the q-GGMRF and the quadratic function should occur. In the following.2. A comprehensive text on this topic is found in [63]. Convex potential functions include the Huber penalty (see e. an approximation of the Huber prior is obtained. [109]. we observe that the slope of the influence function becomes steeper for low contrasts and flatter for high contrasts. Since the sum of two convex functions is a convex function. and c are used to control the shape of the function. Clearly.3) ˜ The derivative V (f ) is called influence function. for p = q = 2. and describes how much different contrast levels should be penalized1 . V (f ) and RV (f ) are linear if and only if V (f ) is a quadratic function. and should satisfy 1 ≤ q ≤ p ≤ 2 and c > 0. the resulting cost function also becomes convex whereby local minima become global minima. Many different non-convex and convex potential functions have been suggested for image reconstruction and restoration. [21]) and the q-generalized Gaussian Markov Random field (q-GGMRF) penalty proposed by Thibault et al. q. As in the previous chapter. 8. a convex potential function has the advantage that when it is used in combination with a convex data term (such as a weighted square norm). ˜ Here. 8. 8. As the value of q is reduced from two towards one. As q decreases from two towards one. we differentiate zV (f ) and multiply with the positive definite matrix V to obtain the update step ˜ ˜ fk+1 = fk − αk V(Q(Pfk − preb ) + β RV (fk )).g. Examples of non-convex potential functions are the saturated quadratic function and the Geman and McClure potential function [29].4) The parameters p. we will focus on the q-GGMRF function V (f ) = |f |p . (8. the potential function becomes increasingly v-shaped. As demonstrated in the right of Fig.2) ˜ RV (f ) = i=1 j=1 dij (vi + vj ) V (fi − fj )ei . 2 (8.5) and is illustrated in Fig. In the left of Fig. the regularization operator RV : RN → RN is given by N N (8. 1 The details of the regularization term are related to the statistical description of an image as a stochastic vector belonging to a Gibbs distribution.8). .

q = 1.0. c = 20 800 600 Penalty 400 200 0 −40 Penalty p = 2.0. 40 Quadratic (q = p = 2) 30 20 Influence (HU) 10 0 −10 −20 −30 −40 −40 −30 −20 −10 0 10 Contrast (HU) 20 30 40 p = 2. c = 20 p = 2.2 Description of the reconstruction method 97 Quadratic (q = p = 2) p = 2. c = 30 −20 0 20 Contrast (HU) 40 −20 0 20 Contrast (HU) 40 Figure 8.0.0.3.0. q = 1.8. q = 1. c = 20 p = 2. c = 10 p = 2.3.3. c = 20 p = 2.1: q-GGMRF potential functions for different values of q and c. q = 1.2: q-GGMRF influence functions for different values of q. q = 1. q = 1.7. c = 20 p = 2. q = 1. c = 20 Figure 8.0.0.0.0.7. c = 20 800 600 400 200 0 −40 Quadratic (q = p = 2) p = 2.0.3. . q = 1.0.3. q = 1.

3. showing difference norms log10 ( fk+1 − fk /C) and SSP full width at half maximum (FWHM) as a function of iteration number. an edge based method yields a different result than a line based method. Assuming this similarity to be valid also for non-linear regularization we limit the experiments in this chapter to the WLS-IFBP method. 60HU. 400HU. for linear regularization. This is demonstrated in Fig.1 Experiments Spatial resolution versus local contrast In the previous chapter.3. In all experiments presented throughout this chapter. One could argue that input data should be pre-processed as little as possible. more iterations are required for convergence. and 500HU. 80HU. which is the traditional choice. Some cross-plane resolution contributions seem to remain. the new update formula becomes fk+1 = fk − αS(Q(Pfk − preb ) + βRV fk ). projection data have been pre-filtered as described in Chapter 5. the results of WLS-IFBP and RIWFBP were found to be very similar. For instance. our choice simplifies comparison with results from previous chapters. (8. As the regularization parameter q is decreased from two towards one. the difference norm is still decreasing after 50 iterations. Furthermore. we have used the solid sphere phantom from previous chapters. 40HU. resolution measurements become structure dependent. Therefore. To assess the spatial resolution for different contrast levels. 200HU. 20HU. .98 Nonlinear regularization For the RIWFBP method. With nonlinear regularization. The different contrast levels studied were 10HU. and that dij should be selected as inverse distances to the center voxel. 300HU. but with different linear attenuation factors. the coefficients dij have been selected so that the nonlinear regularization described here perfectly coincides with the linear regularization from previous chapters when p = q = 2 (βxy = 1 and βz = 1. For low values of q. 150HU.5). However. 100HU.3 8. in the following one should keep in mind that the presented MTFs and SSPs are to be interpreted as characterizations of edge responses. and also to prevent the influence of nonlinear partial volume effects.6) where N N RV (f ) = i=1 j=1 dij V (fi − fj )ei . (8. The fact that MTF and SSP measurements are phantom dependent for nonlinear reconstruction methods is further discussed in the next section. different linear attenuation factors were obtained by linearly scaling the projection data. 8.7) 8. For practical reasons.

6 Quadratic p = 2.4 1. q = 1. where c = 20. 8. By increasing c. for q = 1.2 1 0 10 20 30 40 Iteration number 50 Figure 8. While. c = 20 1. c = 20 1.0. As q decreases. . the ρ50% for the quadratic potential function is constant with respect to the contrast level.6 shows Full Width at Half Maximum (FWHM) values of the SSP as a function of contrast level after 50 iterations.0. q = 1. c = 20 0 log 10( f k+1 − fk ) −1 −2 −3 −4 −5 0 10 20 30 40 Iteration number FWHM (mm) p = 2. Fig. q = 1. In this particular case.7. Here. Hardly any differences can be observed. ρ50% curves are shown for different values of c.4 shows ρ50% values2 as a function of contrast level after 50 iterations.0. Fig. 8. 8. Fig. at which the amplitude has dropped to 50%.3 and c = 20. i. Right: SSP Full Width at Half Maximum (FWHM) as a function of the iteration number. q = 1.4. Since a quadratic potential function corresponds to a linear reconstruction method.5 shows the same plots as Fig. the value of the regularization parameter q is varied. the MTF ρ50% increases with approximately 10%.0. reducing q has a greater impact than it has on in-plane resolution (MTF ρ50% ).0. 8.6 is the unexpected behavior of the SSP FWHM curve for q = 1. the curve drops rapidly from 1. convergence seems to be obtained in approximately ten iterations. 8. c = 20 p = 2.4 but after ten iterations.3: Left: the difference norm log10 ( fk+1 − fk /C) plotted against the iteration number for central images (z = 0mm) of resolution phantom reconstructions with different choices of regularization parameters q. a decrease of more than 30% is observed for the SSP FWHM at contrast levels above 300HU.50mm at 10HU to 2 The ρ50% value of an MTF is the frequency.0.8.3. in terms of in-plane resolution. To the left. Initially.0. To the right in Fig. 8. q = 1.0.7. Another observation in Fig.3 Experiments 99 Quadratic p = 2. c = 20 p = 2. the intersection of the ρ50% curves moves towards higher contrast levels..3. c = 20 p = 2.e. q = 1. the intersection of the ρ50% curves is located at a contrast level of approximately 30HU.0. a slight reduction of ρ50% for low contrasts and a slight increase of ρ50% for high contrasts is observed.

5: ρ50% of the MTF for different contrast levels and regularization parameters q and c. c = 20 q = 1.6 3.4 0 100 200 Contrast (HU) 300 ρ 50% (lp/cm) 4 3.3. c = 20 q = 1. c = 20 q = 1.4: ρ50% of the MTF for different contrast levels and regularization parameters q and c.6 3.0.3. c = 10 q = 1.8 3. Quadratic q = 1. c = 20 q = 1.4 0 100 200 Contrast (HU) 300 ρ 50% (lp/cm) 4 3.100 Nonlinear regularization Quadratic q = 1.6 3. . c = 20 4 ρ 50% (lp/cm) 3. c = 20 q = 1. The number of iterations equals 10.8 3. c = 20 4 ρ 50% (lp/cm) 3.8 3.3. The number of iterations equals 50.3.7. c = 30 100 200 Contrast (HU) 300 Figure 8.7. c = 30 100 200 Contrast (HU) 300 Figure 8.4 0 Quadratic q = 1.0.3.3.8 3.6 3.3.3.4 0 Quadratic q = 1. c = 10 q = 1. c = 20 q = 1.

This is verified in Fig.8. c = 20.8 0 100 200 Contrast (HU) 300 1. 8. Derivations of edge based measurement methods rely on the linearity to show how the PSF can be calculated from an edge response.9 0. c = 20 q = 1. 8.3. and levels off at approximately 1.2 Edge versus surface based resolution measurements Most expressions used for characterizing spatial resolution in an imaging system are based on the assumption that the system is approximately linear. Under this assumption.2 1.2 1.0. For q = 1. For a contrast level of 200HU and q = 1. the number of iterations needed for convergence of cross-plane resolution is high for values of q close to one.0 mm (not visible in the figure). c = 20 q = 1. c = 30 100 200 Contrast (HU) 300 Figure 8. the spatial resolution properties are fully characterized by the spatially variant or invariant point spread function (PSF).4 FWHM(mm) FWHM(mm) 1. The number of iterations equals 50.6: SSP FWHM for different contrast levels and regularization parameters q and c.1 1 0. 8.9 0.8 0 Quadratic q = 1.8.1 1 0.0.3. Although the plots are similar in both cases.6 but for 50 iterations.4 1. SSPs for this case are shown in Fig.3.3. the improvements are less pronounced after 10 iterations than after 50.7. c = 20 q = 1. the relative amount of low contrast smoothing becomes smaller. the previously observed 30% decrease of FWHM after 50 iterations is approximately 20% after 10 iterations.3. As shown already in Fig. and the SSP approaches that of the non-regularized method. As the contrast level increases to 500HU. showing the same plots as Fig.3.3 1. the FWHM starts to increase. As the contrast further increases.7.3 1. around 0.3. c = 20 1. it seems like the combination of much smoothing for low contrasts and very little smoothing at high contrasts helps to sharpen the SSP more than in the non-regularized case.8mm at 200HU.3 Experiments 101 Quadratic q = 1. c = 10 q = 1. 8. . 8.

c = 20 1.3.3.2 1.3.2 1. the number of iterations equals fifty.1 1 0 100 200 Contrast (HU) 300 FWHM(mm) q = 1.5 1. In all cases.7.2 0.8 2 Figure 8.0.4 1. . The number of iterations equals 10.7: Comparison of SSPs for quadratic regularization. and strongly nonlinear regularization (q = 1.2 1.3 1.0.4 0.8 1 1.4 FWHM(mm) 1.5 1.4 Quadratic No regularization q = 1.2 0 0 0. c = 20.102 Nonlinear regularization 1 0. c = 10 q = 1. Quadratic q = 1.1 1 0 Quadratic q = 1. no regularization. c = 20 q = 1. c = 20 q = 1.6 1.3.8 0.4 Distance from slice center (mm) 1. c = 20 1.0. c = 20) for the two contrast levels 200HU and 500HU. c = 20.3 1.6 0.6 SSP 0. c = 30 100 200 Contrast (HU) 300 Figure 8. 200HU q = 1. 500HU 0.8: SSP FWHM for different contrast levels and regularization parameters q and c.0.

Surface 0.8 1 1.2 1. For this contrast level.3. Fig.4 0. .2 0.2 0 0 0. c = 20) reconstructions of the two phantoms described in Section 4. the edge FWHM is reduced by 17% and the line FWHM is reduced by 12%. q = 1. 8. the SSP measurements differ very little from each other. Surface f 10. The number of iterations equal ten. Therefore.4 Distance from slice(mm) 1. and consequently. the surface SSP is clearly wider than the edge SSP.8 2 Figure 8.3.4 0. and the other one is a 0. q = 1.0. p = 2. One of these is a solid sphere referred to as the edge phantom.3. The non-normalized peak of the line SSP is 80HU.6 1.6 SSP 0. for the nonlinear reconstructions. However.3 Experiments 103 1 0. An imaging system using nonlinear regularization as presented in this chapter cannot be assumed to be “approximately linear”. This difference is explained by (i) thickness dependent dampening of the reconstructed profile of the thin shell phantom (0. in the previous section it was pointed out that the presented MTFs and SSPs should be interpreted as characterizations of edge responses.3. q = 1.8 0.1mm⇒ 1000HU→ 80HU).9: Comparison between edge based and surface based SSP measurements in presence of nonlinear regularization.9 shows SSPs for WFBP and WLS-IFBP(f10 .6 WFBP.1mm thick shell referred to as the surface phantom. the measured spatial resolution for such a system depends on the type of structure studied. and (ii) the difference in shape between the profiles of the shell and edge phantoms. Both phantoms have a density corresponding to 1000HU above air. Edge WFBP.8. Edge f 10. For the WFBP reconstructions. We illustrate the dependence of structure with an example.

the pixel noise is inversely proportional to the square root of the number of photons.3 Artifacts and noise Fig. where the nonlinear and linear methods give similar results. and partial volume artifacts originating from the high contrast structures in the inner ear [129]. an effect which seems rather independent of the parameter c. The nonlinear WLS-IFBP in Fig. Here. The image is corrupted by streak artifacts plus a combination of cone-artifacts. To the left. Fig. attained with nonlinear regularization (p = 2. the linear WLS-IFBP method gives a better result with less pronounced wind-mill artifacts. To examine the noise properties for the nonlinearly regularized WLS-IFBP. In addition. 8.104 Nonlinear regularization 8. and for different values of the regularization parameter c. Depending on the choice of regularization parameter c. but are still visible in this particular greyscale window. In Fig. the nonlinearly regularized WLS-IFBP suppresses noise at low noise levels.10 d).13 shows three axial images of the physical head phantom reconstructed with WFBP. after ten iterations with WLS-IFBP using quadratic regularization. σnq are noise measurements for the quadratically regularized WLS-IFBP method. we have studied head phantom reconstructions for different numbers of non-attenuated photons. i.. In relation to the linearly regularized WLS-IFBP method.3. For any linear reconstruction method. for nonlinear methods. as well as improves the cross-plane resolution as can be readily seen in the encircled areas. i. The reduction is slightly higher for the nonlinearly regularized reconstructions. As shown in Fig. can be controlled by changing the regularization parameter c.13 b). 8. this σe value is reduced from 102HU for the quadratic (linear) case to somewhere between 95HU and 96HU. wind-mill artifacts.10 b) shows an axial image of a WFBP head phantom reconstruction from noise-free data. and non-linear WLS-IFBP. there is a reduction of this value from 10. this does no longer hold true. These observations are confirmed by the σe measurements plotted in Fig. 8.12 shows σn /σnq for noise levels corresponding to {4k × 106 }2 k=−2 nonattenuated photons. Fig. 8. 8. 8.e.4HU to just below 9HU. To measure the maximum noise levels. The WFBP reconstruction is corrupted by windmill artifacts. The transition point. linear WLS-IFBP. measurements have been restricted to the center of the reconstructed objects. increasing the number of photons with a factor of four results in a reduction in pixel noise by a factor of two [46].3.13 c) further reduces the amplitude of these artifacts. the artifacts are even more alleviated. . and amplifies noise at high noise levels.e. for which the above presented proportionality holds. The right part of this figure shows σe calculated on low contrast regions. the strong artifacts originating from the inner ear have been alleviated..10 c). the nonlinear regularization also improve the image in terms of spatial resolution. 8. q = 1. However. c = 20). 8.0. the measurements stem from the whole image. As shown in Fig.11. For all iterative methods.

8.3 Experiments 105 a) Sampled phantom b) WFBP reconstruction c) WLS-IFBP.3.10: Axial images of noise-free head phantom reconstructions. f10 . f10 . Greyscale window:−25HU. q = 1. . c = 20 Figure 8.25HU.0. p = 2. p = q = 2 d) WLS-IFBP.

2 1. . σnq = 22.106 Nonlinear regularization p = 2. c c c c 1. the number of iterations are 10.95 0. c = 20 5 Iteration number 10 Figure 8.0 q = 1.3.1 σn /σnq 1.5 9 8.3.3.11: WLS-IFBP σe values measured on head phantom reconstructions. = 1.5 8 0 q = 1. low contrast regions (HU) 10. whole image (HU) 101 100 99 98 97 96 95 0 5 Iteration number 10 q = 1. In all cases. c = 40 RMSE.12: σn /σnq against the number of unattenuated photons.0.3. c = 30 102 RMSE. σnq are noise measurements for WLS-IFBP with quadratic regularization. = 1.4HU = 10 = 20 = 30 = 40 10 10 Number of unattenuated photons 6 10 7 Figure 8.6HU 2. c = 10 q = 1. = 1.3.15 1.9 5 11.3.3.8HU q q q q = 1.3.5 10 9. q = 2.1HU 5. The noise measurements were made on central parts of the phantom.05 1 0.6HU 1.

13: Axial images of head phantom reconstructions. . p = q = 2.3. q = 1. f10 .0 c) WLS-IFBP. f10 .3 Experiments 107 a) WFBP reconstruction b) WLS-IFBP. Greyscale window ±75HU. c = 20 Figure 8.8.

Fig.108 Nonlinear regularization Fig. resulting in sharpening of high contrast structures and softening of low contrast structures.14 shows axial images of physical thorax phantom reconstructions. Fig. most improvement take place within the first ten iterations. we observe the presence of wind-mill artifacts in the WFBP reconstruction. Another way of describing this is to view the nonlinearly regularized reconstruction as using two reconstruction kernels3 : one sharp kernel for high contrast structures.1). Again.9.14 c).g. As illustrated by the simple one-dimensional example in Fig. 8. for potential functions that deviate much from a quadratic function. However. Such step lengths can be found by performing a so called line search in each iteration.15. A convergent method is obtained if the step lengths αk satisfy the so called Wolfe conditions [71]. 8. In terms of artifact and spatial resolution.7 and the low-pass filtering S introduced in Section 5. and relatively high penalties for low contrast structures. A practical line search method 3 A reconstruction kernel is a rampfilter convolved with a resolution modifying filter as described in Chapter 5. two values of β were used.14 d) than Fig. and one soft kernel for low contrast structures. As shown in Fig. the nonlinear WLS-IFBP with β = 0. for q close to one.4 Discussions and conclusions The results in this chapter indicate that nonlinear regularization can be used for improving results of WLS-IFBP reconstruction with respect to wind-mill artifact reduction and high contrast spatial resolution.1. For the non-linear WLS-IFBP method. by increasing β to 0. 8. 8.6 does not improve the result much over the linear WLS-IFBP. further suggests that the convergence rate could be improved by using larger step lengths for high contrast structures than in low contrast regions.15 illustrates a problem associated with step lengths that are constant with respect to iteration number. the optimal step length α depends on the shape of the cost function. it is still better than for the linear methods. 8.15. Although the high contrast spatial resolution is worse in Fig.14 c). . and how the amplitude of these to a certain extent can be reduced with the linear WLS-IFBP method. While this works fine for a quadratic potential function (see Section 5. Another possibility for improving the convergence rate is to use a more advanced optimization method such as the preconditioned conjugate gradient method or ordered subsets acceleration. 8. e. This is partially explained by the constant step length α = 0. without increasing noise in low contrast regions. an infinite cycle of sub-optimums might occur for a nonquadratic potential function. further artifact reduction can be obtained. However. more iterations are needed until no further improvement can be observed. as well as preliminary experiments. 8. 8. This is an effect of using relatively low penalties for high contrast structures.

14: Axial images of thorax phantom reconstructions. p = q = 2. q = 1. β = 1. f10 . . β = 0.6 d) WLS-IFBP.3.8. Greyscale window: ±150HU.0 c) WLS-IFBP. q = 1. c = 50. f10 . c = 50.4 Discussions and conclusions 109 a) WFBP reconstruction b) WLS-IFBP.0 Figure 8.3. f10 .

requiring no forward projections or backprojections has been proposed by Fessler and Booth [25]. and c = 20. α = 0. 1+|x/c| p−q 50 α = 0. In all cases. the noise estimation problem is avoided by downweighting noisy data contributions relative to the regularization term. leading to blurring along directions orthogonal to the ray projection. From a clinical point of view. In order not to enhance noise. q = 1. This is an important difference between the methods studied here and statistical methods. The parameters for f2 are p = 2. the maximum noise level must be estimated and used for properly choosing the regularization parameter c.110 Nonlinear regularization f1 (x) = x2 /2. The q-GGMRF potential function adds a number of regularization parameters to the cost function.15: One-dimensional example of gradient descent optimization of two functions f1 (x) and f2 (x) using two different constant step lengths α.9 1500 1500 f2 (x) = |x| p .8 50 1000 1000 500 0 0 −50 f2 (x) = 0 x |x| p . In statistical reconstruction methods. α = 1. artifact suppression and improved spatial resolu- . the iterative method has been initialized with x0 = −50.9 f1 (x) 500 f2 (x) −50 0 x f1 (x) = x2 /2. These parameters must be carefully adjusted for every type of examination that the CT system is intended for.8 1500 1500 f1 (x) 500 f2 (x) −50 0 x 50 1000 1000 500 0 0 −50 0 x 50 Figure 8. 1+|x/c| p−q α = 1.

8. Borsdorf et al. it remains unclear whether the observed noise reduction in low contrast regions improves detection of low contrast objects. However. A human observer study is needed to draw any further conclusions in this matter.4 Discussions and conclusions 111 tion of the nonlinearly regularized WLS-IFBP method will highlight low contrast structures otherwise distorted by artifacts or located close to bone.g. [4]). . (see e.

112 Nonlinear regularization .

we must allow the basis functions to be view-dependent as illustrated in Fig. which is a rather poor model of the physical projection event. The fact that the gantry is rotating during acquisition is also ignored. (9.3. Still.1 that a measurement pi acquired from a reconstruction result f can be expressed as pi = wi (r)fc (r)dr. In the following section.9. sim1 We use the words “single ray” to emphasize that the Joseph method appearing in previous chapters calculates one single line integral per measurement. 3. since the Joseph method generates projection data that are very similar to the method by K¨hler and Turbell [51]. We recall from Section 3. The functions wi : R3 → R are so called irradiation functions that define how fc contributes to each of the measurements pi respectively. For the single ray Joseph method.1 Irradiation functions As discussed in Chapter 3. that employs a view-independent tri-linear basis function o [110]. 2 To make (9. throughout previous chapters we have employed the single ray Joseph method1 .1) R3 where fc : R3 → R is the continuous function obtained by interpolation with the basis function b : R3 → R2 . and can be used to incorporate accurate models of the data acquisition in the forward projection operator P. we investigate the effects of improving the projection operator P by using irradiation functions consisting of multiple Dirac lines. . This should not have dramatic effects on the reconstruction results. the irradiation function wi for a certain measurement pi is a single Dirac line intersecting the source and detector of this specific measurement.1) applicable to Joseph forward projection. This is equivalent to assuming that the source and detector are point shaped. iterative reconstruction offers the possibility to accurately model the data acquisition process in the forward projection P.Chapter 9 Data acquisition modeling 9.

2) − cos α sin η. In the same way as for the source. We start with a description of the improved cone-beam geometry model. . and rSC ± (lS /2)ˆSL . The four .1. presumably caused by complete suppression of certain high frequencies in the rebinning. T − sin α. By formulating a rampfilter weighted least squares problem in the cone-beam data domain. the input data should be preprocessed as little as possible. previous experiments in [102] have shown that putting the rebinning inside the iterative loop leads to strong overshoots. 0.2. sin(α + β). However.3) Let dβ and dq be factors describing the relation between the detector sampling distances and the actual width and height of the detector elements. T αP T 2π . Instead.1 and 9. the detector elements are aligned to the z-axis as shown in Fig. (9. and focus on improving the reconstruction by modeling the low-pass filtering in the acquisition process. 1 T − sin(α + β). and then explain how this model can be applied to the rebinned geometry. 9. 0. and the unit vectors ˆSW and ˆSL pointing along the main directions of r the rectangle. rebinning should not be a preprocessing step but rather be put inside the iterative loop.. i. as well as gantry rotation.2. When projected onto the xy-plane. 9. The center rSC of the rectangular r source. the detector elements are described by rDC (α. 9. 0 T qh 2RS T − cos(α + β). We assume that the source (the focal spot) is shaped like a rectangle sticking to the coneshaped anode as shown in Fig. Any point r on the rectangle can now be expressed as a convex combination of the four vectors r r defining the corners of the rectangle: rSC ± (wS /2)ˆSW . Furthermore. the detector elements face the source as shown in Fig. cos α. but calculating strip integrals instead of line integrals. cos η The so called anode angle η is the angle between ˆSL and the z-axis. RS sin α. we accept the low-pass filtering caused by rebinning.e. are given by rSC (α) ˆSW (α) r ˆSL (α) r = = = RS cos α. 9.2. β. − sin α sin η.2 9. (9. We use a rectangular model also for the detector elements. q) = ˆDW (α. or entirely avoided. 0 . and . it could be possible to avoid rebinning entirely.1 Multiple ray forward projection An improved acquisition model To get most out of any modeling attempt. β) r ˆDH r = = rSC (α) + RSD − cos(α + β). This is followed by a presentation of a projection method based on the Joseph method. but this has not been investigated so far.114 Data acquisition modeling ulating the spatial extent of the source and detector.

and involved unit vectors in the xy-plane.2: Illustration of the shape of the source and detector. Note that the sizes of the source and detectors are exaggerated. Note that the sizes of the source and detectors are exaggerated. .9.2 Multiple ray forward projection 115             Figure 9.1: Illustration of the shape of the source and detector. and involved unit vectors in the xz-plane.         Figure 9.

.. The arrangements of the points on the source and detector elements are illustrated in Fig. Let NSW and NSL be the number of source points along the ˆSW and ˆSL respectively. vectors pointing out the corners of a detector element are then given by rDC ± r r (wD /2)ˆDW and rDC ± (hD /2)ˆDH . NSW − 1. NSL − 1. NDH − 1.3. and hD = RSD dq Δq h .. 2RS (9. ...6) m = 0. NDW − 1..3: Arrangement of points for a multiple ray forward projection model. the next step is to divide these rectangles into multiple points. . l] = rSC (αi )+ + wS NSW − NSW − 1 lS + k ˆSW + r 2 NSL − NSL − 1 + l ˆSL . 9.. and to calculate measurements as mean values of line integrals between all source points and all detector points. the points used for simulating the spatial extent of the detector are given by rD [i.5) k = 0..4) Having defined the rectangles for the source and detector. n = 0. . l = 0. k. . m. r 2 (9.. .116 Data acquisition modeling         Figure 9. n] = rDC (αi . Similarly. r 2 (9. r r The points used for simulating the spatial extent of the source are then given by rS [i. where the width wD and height hD are given by wD = RSD dβ Δβ . βi .. qi )+ + wD NDW − hD NDW − 1 + m ˆDW + r 2 NDH − NDH − 1 + n ˆDH ...

the SSP FWHM remains unchanged at 1. input data have been generated as described in Chapter 4. Interestingly. The WFBP reconstruction in 3 The “standard backprojector” B refers to a backprojector employing linear/bilinear interS polation adapted to the detector sampling grid. since the total cost of a multiplication with P is directly proportional to the product NSL × NSW × NDH × NDW × NGR . using the scanning parameters in Table 4. the more accurate the model becomes. Unfortunately.4.53mm to 1. 9.1.3). as well as azimuthal resolution. If the value of NDH is increased to four. and ˆDH are computed correctly. ˆDW .p) by using the rebinning equations β = sin−1 (p/RS ) and α = θ − β.2) and (9. DC-aliasing artifacts during backprojection can be reduced by increasing the number of rays per measurement using the techniques described in the previous section. For more accurate modeling of any of the previously presented entities (e.β) from (θ. we present results obtained not only with the “standard backprojector”3 BS . k. Two resolution phantoms were used: one located in the iso-center and one located 150mm off-center for measuring off-center radial resolution. NSL ) should be set to a value larger than one. due to the mismatched size of the interpolation function.29mm. we need to make sure that the unit vectors ˆSW . 9.7) Using the same technique as for the source and detector modeling.9. When applying this model in the rebinned geometry. NDH . . and applying (9.2 Experiments We have studied spatial resolution for different choices of the forward projection parameters NSL .29mm. This is illustrated in Fig. For example. NDW . m. This is done r r r r by calculating (α. NSW . In the experiments.5. In contrast to previous chapters. ˆSL . In all experiments. We use the word “standard” because this is the backprojector most frequently found in the literature. the ith measurement can be expressed as pi = 1 NSW NSL NDW NDH NSW −1 NSL −1 NDW −1 NDH −1 fc (r)dl. the acquisition model can be further improved by taking the gantry rotation into account. and NGR . length of the source).g. The higher the N -value is. 9. As illustrated in Fig.2 Multiple ray forward projection 117 With Γiklmn denoting the line intersecting the points rS [i. n]. the N -values have been empirically selected such that any further increase in N -value yields an improvement that is very small in relation to the already obtained improvement.2. its corresponding N -value (e. k=0 l=0 m=0 n=0 Γiklmn (9. l] and rD [i.g. DC-aliasing artifacts may occur when using PT corresponding to the Joseph method with only one ray per measurement. high N -values quickly leads to prohibitively high computational costs. but also with the transpose PT of the forward projector as suggested by the derivation of the gradient descent method in Section 7.2. NDH = 3 reduces the SSP FWHM from 1. We let NGR denote the number of sub-angles used to simulate this effect.

118 Data acquisition modeling                   Figure 9. 9. These DC-aliasing artifacts occur mainly because the interpolation function in the z-direction is too small: the detector sampling distance is hΔq /2 = 1. and the size of the interpolation function is determined only with respect to the much smaller voxel volume sampling distance Δz = 0. false high frequencies occur in the backprojected images. . and how it is affected by modeling of the source length NSL and detector element height NDH in the forward projection model. the size of the interpolation function becomes more adapted to the detector sampling distance.5 b) we see that WLS-IFBP with BS as backprojector does add some high frequency artifacts. Since the interpolation function in b) is not adapted to the projection data sampling distance. However. 9.5 a) is completely free from artifacts originating from interpolations in the backprojection. In Fig. Fig.5mm. the amplitude of these artifacts is low. and the amplitude of the DC-aliasing artifacts is reduced as shown in Fig. NDW . Fig. and NGR have negligible impact on the cross-plane resolution.2mm.5 c). We begin with studying cross-plane resolution in the iso-center. It is reasonable to assume that the other parameters NSW .4: Illustration of DC-aliasing during backprojection.5 d). 9. modeling of the detector element height with NDH = 3 will always be used for suppressing DC-aliasing artifacts. 9.5 e) and f) show that this type of artifacts can be further reduced by adding more rays corresponding to the detector width and the spatial extent of the source. severe artifacts occur as shown in Fig. 9. and can be further reduced by increasing the sampling density of the reconstruction volume [102]. As the number of Dirac lines is increased from one to three in the detector height direction. In the following experiments. When BS is replaced with PT .

Greyscale window (-1025HU.2 Multiple ray forward projection 119 a) WFBP reconstruction b) WLS-IFBP. PT NSL = 1. NDW = 2 Figure 9. NDW = 2 f) WLS-IFBP. f10 . NDW = 1 d) WLS-IFBP. f10 . f10 . PT NSL = 1. PT NSL = 1. NSW = 1 NDH = 1. PT NSL = 3. In all cases. 975HU). . NSW = 1 NDH = 3. f10 .9. BS NSL = 1. the NGR equals one. NDW = 1 e) WLS-IFBP.5: Axial images of MTF phantom reconstructions for different backprojectors and number of rays per measurement. NDW = 1 c) WLS-IFBP. NSW = 1 NDH = 3. NSW = 1 NDH = 1. f10 . NSW = 2 NDH = 3.

120 1. To see how the source length affects the result. Both BS and PT were used as backprojectors. 9. 9. but decays more rapidly for high frequencies. NDW = 2).7 shows the SSPs from Fig. the resolution does not continue to improve after the ten first iterations. and modeling of both source and detector element widths (NSW = 2. in contrast to the cross-plane resolution case. these parameters have all been set to one.25 0 5 10 Iteration number Data acquisition modeling BS . consider . In-plane resolution can be divided into radial resolution in the iso-center. Although the MTFs differ very little. Fig. Interestingly. PT . the SSP FWHM stops improving after approximately 10 iterations. 9. higher spatial resolution is obtained with BS than with PT . the FWHM has decreased from 1.55 SSP FWHM (mm) 1. the MTF for PT follows the MTF for BS for low frequencies. and is obtained with modeling of both detector and source width (NSW = 2. Fig. as well as increasing NSL from one to three lead to an amplification of the overshoots. and radial and azimuthal (angular) off-center resolution. Unexpectedly. At this point. PT . but also by the gantry rotation (mainly azimuthal resolution) and source length (mainly radial resolution). In addition to the previous observations.3 1. As seen in Fig. Except for NSL . NS L = 1 NS L = 3 NS L = 1 NS L = 3 15 Figure 9. The off-center resolution is affected not only by the source and detector widths. we see that using BS instead of PT . BS . The resolution in the isocenter should be affected only by the widths of the source and detector elements. Fig.6 1.8 left shows MTF ρ50% measurements for modeling of only detector element width (NSW = 1. 9.45 1. and 1. the detector height is taken into account with NDH = 3.5 1. NDW = 2). and the standard backprojector BS . In all plots.4 1. one interesting observation is that by modeling both source and detector element widths. NDW = 2). Similarly to the cross-plane resolution case.8 shows MTFs after ten iterations.35 1. in the case of the PT backprojector.43mm with BS .6: SSP FWHM curves as a function of iteration number for WLS-IFBP.53mm to around 1.6 after ten iterations. other multiple ray parameters equal one. Therefore.6.29mm with PT . 9. The highest increase in MTF ρ50% is approximately 10%. increasing NSL from one to three leads to a slightly increased FWHM. Only small differences are observed when NSL is increased from one to three.

the detector width and height are taken into account with NDH = 3 and NDW = 2.5 Distance to center (mm) 3 Figure 9. N S L = 3 1 0.8 3. BS .9.7 3.2 0 0 0. N S L = 1 BS . other multiple ray parameters equal one. Right: Blow-up of detail.5 2 2.8 0.05 0.8 SSP SSP 0. N S W = 1 PT . the detector height is taken into account with NDH = 3. Both BS and PT were used as backprojectors. Left: complete SSP.4 0. other multiple ray parameters equal one.5 1 1. N S L = 1 PT . .1 PT . N S W = 2 4 6 8 Frequency (lp/cm) 10 Figure 9. N S W = 1 BS .05 0.2 Multiple ray forward projection 121 BS .7: SSPs for WLS-IFBP after ten iterations. N S L = 3 0 1 1. 10 iterations 0. In all plots. N S W = 2 PT .6 0.5 Distance to center (mm) 2 −0. In all plots.8: Left: iso-center MTF ρ50% values for WLS-IFBP.2 0 0 2 BS .4 0. Right: MTFs after ten iterations.6 0. N S W = 1 PT . N S W = 2 1 3. Except for NSW . Both BS and PT were used as backprojectors. Except for NSL .9 MTF ρ 50% (lp/cm) 3.6 0 5 10 Iteration number 15 MTF. N S W = 2 PT . N S W = 1 BS .

N S W = 2. i. Gantry rotation should not affect the off-center radial resolution.9: Left: off-center radial MTF ρ50% values for WLS-IFBP.122 Data acquisition modeling BS .2 0 0 2 BS . 9. Both BS and PT were used as backprojectors. the improvements of ρ50% values increase to 33% for BS and 22% for PT . Ignoring the width of the source. With gantry modeling.8 0.2 3 2. N S L = 1 PT . an object located 150mm off-center. N S W = 1. The left of Fig. corresponding numbers are 6. This object is irradiated at a maximum fanangle of β = sin−1 (150mm/RS ). N S L = 1 PT .e. the ρ50% values increase with 3. at this fan-angle the object “sees” the source as if it is lS × 150mm/RS ≈ 3mm wide. As shown in the MTF plots to the right.6 0 5 10 Iteration number 15 MTF. N S W = 2. N S W = 2. N S W = 1. N S L = 1 BS . N S L = 3 4 6 8 Frequency (lp/cm) 10 Figure 9. the MTF drops off faster with PT than with BS . Only modeling the spatial extent of the detector results in an increase in ρ50% of approximately 3%. N S W = 1. Without gantry rotation modeling. the width of the resulting x-ray beam will cause blurring in the radial direction. independent of the backprojector type. Due to the large number of rays involved.5% for PT and 5. N S W = 2. Clearly. as well as backprojector type.9 shows how off-center radial MTF ρ50% values are affected by modeling of the detector elements and source. In all plots. N S L = 3 PT . 9. the detector element widths and heights are modeled with NDW = 2 and NDH = 3. N S W = 1. many of the models presented here are computationally too expensive to be applied to any full size clinical data set. We also note that with modeling of source width and height.8 2.6 MTF ρ 50% (lp/cm) 3. and has not been modeled here.4 0. When also the length and width of the source are taken into account. Right: MTFs after ten iterations. Since modeling of the detector height and width gives an easily observable increase .10 shows off-center azimuthal MTF ρ50% values and MTFs with and without gantry rotation modeling. 10 iterations 1 0.6 0. N S L = 1 BS . NGR = 1. N S L = 3 3.0% and 10% respectively. Fig. the improved modeling does not change the location of the first zero in the MTF.4 3.2% for BS . N S L = 3 PT .

the detector element geometry has been modeled with NDW = 2 and NDH = 3.11 shows reconstructed images of the physical head phantom with linear regularization. accurate modeling of the acquisition process can improve image reconstruction in terms of spatial resolution. 9. Right: MTFs after ten iterations.2 0 0 2 BS .e. 9. 10 iterations 1 0. shows axial images obtained with linear and non-linear regularization. N GR = 2 PT . Fig. in cross-plane resolution at a reasonable cost increase. Unfortunately. The PT result has a slightly higher cross-plane resolution.6 0.10: Left: off-center azimuthal MTF ρ50% values for WLS-IFBP. N GR = 2 0 5 10 Iteration number 15 4 6 8 Frequency (lp/cm) 10 Figure 9. N GR = 2 PT . Both BS and PT were used as backprojectors. The amount of improvement depends .12 d) shows the result obtained with PT . the detector element widths and heights are modeled with NDW = 2 and NDH = 3. i. we see that the high contrast cross-plane resolution is further improved by the detector element modeling.12 shows how nonlinear regularization can partially be used to control the increased amount of wind-mill artifacts.8 MTF. but as indicated by the dashed circles. We note that the detector element modeling improves the cross-plane resolution (indicated with solid circles). The upper row. 9.3 Discussions and conclusions Clearly. and Fig. N GR = 1 BS .12 a) and b). 9. N GR = 1 BS . 9.12 c) shows the result obtained with BS . another obvious and immediate consequence of the detector element modeling is an increased amount of wind-mill artifacts. N GR = 1 PT .9 2. N GR = 2 3. In all plots. we choose this model for a visual inspection of physical head phantom reconstructions. Fig. 9. N GR = 1 PT . Fig.2. In the lower row. but without any modeling of the spatial extent of the detector elements.2 MTF ρ 50% (lp/cm) 3.2 Multiple ray forward projection 123 BS .4 0. In the area indicated by the solid circles. Fig. The source length and width are modeled with NSL = 3 and NSW = 2. this image also contains more wind-mill artifacts.9.8 0.1 3 2.

9.13 and studying the fixed point f∞ = (BWP + βR)−1 BWp. Δt = 0. if the voxel density is higher than the ray density. i. NDW = 1 b) Axial image.47. In both cases cases.11: WLS-IFBP f10 reconstructions of the physical head phantom using linear regularization. BS NDH = 1.13 c) and d).. and on the kind of backprojector being used. corresponding to the first zero in the amplitude spectrum of the backprojection interpolation function. Δx < Δt or Δz < Δq h/2. Of special interest is the lowest frequency in N (BW). PT gives higher cross-plane resolution than BS . on which parts of the acquisition system that are modeled. when Δt < Δx or Δq h/2 < Δz .4mm) and the crossplane ray density is relatively low (Δz = 0. First we note that input data frequency components belonging to the null space N (BW) cannot be reconstructed. The difference in spatial resolution arising from using BS and PT as backprojectors can be understood by looking at Fig.5mm. Equivalently. β = 0.e. this frequency is higher for PT than for BS . One recurring observation in the previous section is that. 9.13 a) and b). Reconstruction parameters: Q = 0.2mm).e.4. i.5mm. Vice versa. this frequency is higher for . NSW = NSL = NGR = 1.124 Data acquisition modeling a) Axial image. when the ray density is high in relation to the voxel density. NDW = 2 Figure 9. βz = 1. if the ray density is higher than the voxel density as in Fig. βxy = 1. In our scanner configuration. where B equals either BS or PT . 9. Consequently. and BS gives higher in-plane resolution than PT . Δq h/2 = 1. Greyscale window (40 ± 75)HU. the in-plane ray density is relatively high (Δx = 0. higher resolution is obtained with BS than with PT as backprojector. higher resolution is obtained with PT than with BS .0. PT NDH = 3. If the voxel density is higher than the ray density as in Fig.7.

9.2 Multiple ray forward projection

125

a) BS , p = 2, q = 2, c = 20, β = 0.47 NDW = 1, NDH = 1

b) BS , p = 2, q = 1.3, c = 20, β = 0.47 NDW = 1, NDH = 1

c) BS , p = 2, q = 1.3, c = 30, β = 0.8 NDW = 2, NDH = 3

d) PT , p = 2, q = 1.3, c = 30, β = 0.8 NDW = 2, NDH = 3

Figure 9.12: Axial images of WLS-IFBP f10 reconstructions using linear and nonlinear regularization. Greyscale window (40 ± 75)HU. In all cases, NSW = NSL = NGR = 1.

126

Data acquisition modeling

1.2 1 0.8 0.6 0.4 0.2 0 −1 −0.5 0 0.5 Distance (mm)

BS PT

1.2 1 0.8 0.6 0.4 0.2 0 1 −5 0 Frequency (1/mm)

BS PT

5

a) Spatial domain, in-plane
1.2 1 0.8 0.6 0.4 0.2 0 −1.5 −1 −0.5 0 0.5 Distance (mm) 1 1.5 BS PT

b) Fourier domain, in-plane
1.2 1 0.8 0.6 0.4 0.2 0 −0.2 −2 −1 0 1 Frequency (1/mm) 2 BS PT

c) Spatial domain, cross-plane

d) Fourier domain, cross-plane

Figure 9.13: Interpolation functions and their Fourier transforms for BS and PT in the t-direction and q-direction respectively. BS than for PT . Another reason is that the balance between the regularization contribution and data contribution is affected by the amount of low-pass filtering done by the interpolation in the backprojection. A backprojector that preserves more high frequencies results in a sharper iterative reconstruction result. Since the computational cost rapidly increases with the accuracy of the acquisition model, only very simple multiple ray models can be used on full size clinical data sets. Already detector element modeling with NDW = 2 and NDH = 3 results in a forward projection that is six times more expensive than with a single ray model. Therefore, it does not seem reasonable to further improve the model with source and gantry rotation modeling unless there are special requirements on off-center resolution. An interesting alternative to the multiple ray experiments made here is to analytically derive interpolation schemes corresponding to geometrical models of

9.3 A strip integral projection method

127

the source, detector elements, and gantry rotation [14, 15]. This could result in computationally cheaper ways to achieve accurate linear models of the acquisition system. One problem with these methods is that it is difficult to take the exact geometry into account, e.g. the tilted anode in Fig. 9.2. A very simplified variant of this is kind of methods is found in the next section. As previously discussed, the in-plane resolution is limited by initial rebinning of input data. One possibility to further improve in-plane resolution could be to entirely avoid rebinning by performing forward projection and backprojection directly in the cone-beam geometry, and use a rampfilter derived for fan-beam reconstruction. Since normalization would not be done correctly, the first iterates would be severely corrupted by artifacts. An interesting future investigation would be to examine how well these artifacts are corrected by the iterative loop, and how the in-plane resolution would be affected.

9.3
9.3.1

A strip integral projection method
Method description

In this section, we present a forward projection method P that avoids DC-aliasing in its corresponding backprojection PT by calculating strip integrals instead of line integrals. The method was first presented in [104]. Similar methods for have previously been presented by Hanson and Wecksung [33], Ohyama et al. [76], Mueller et al. [68], Ziegler et al. [124, 125], and De Man and Basu [17]. For simplicity, the method is presented in the context of 2-D parallel beam CT. Except for some small details, generalization to 3-D cone-beam or semiparallel geometry is straightforward. Let θ be the projection angle and let t denote parallel displacement as described in Section 2.1. Then, the direction of T the rays is given by s(θ) = − cos θ, − sin θ and the displacement is given by t(θ, t) = t sin θ, expressed as − cos θ
T

. By putting s and t into (3.13), the footprint can be

g(θ, t) = g(s(θ), t(θ, t)) =

1 Δ

b(t(θ, t) + ls(θ))dl.
R

(9.8)

This footprint can be used to construct a line integrating projection operator by, for each voxel, projecting its footprint onto the detector, and sample the projected footprint in the points of the detector elements. Hanson and Wecksung [33] showed how the footprint can be used for calculation t ˜ ˜ of strip integrals. By storing the integral I(θ, t) = −∞ g(θ, t)dt, a strip integral

e.j ± Δt /2) is equivalent to sampling the footprint convolved with a rectangle function of width Δt in the point ti − ti. Similarly to the .i (t) = gJ (θi .128 Data acquisition modeling defined by the displacements t1 and t2 of the basis function b can be calculated as t2 t2 b(t(θ. t)dt = Δ(I(θ. t2 ) − I(θ. gJ. By combining the Joseph method with the strip integration technique. The Fourier transform of hi (t) is given by Hi (ρ) = Δ2 sinc2 (Δx ζ(θi )ρ) sinc(Δt ρ).j = xj sin θi − yj cos θi . 1−|t|).9).j be the projection of pixel j onto the detector along the direction corresponding to the ith measurement. the strip integral method is well suited for an efficient computer implementation. Δt =:hi (9. is based on the Joseph method [41] in the way that it employs the same footprint function gJ (θ.12) In order to avoid DC-aliasing during backprojection. Let θi and ti be the parameters defining the ray corresponding to measurement i. which is obviously accomplished by the sinc factor. The function hi describes how the pixels of the image contribute to the measurement i during forward projection. | sin θ|).9) =Δ −∞ g(θ. Hi (ρ) must be zero for ρ = kΔ−1 .11) Here. t)dt = g(θ. To see this. and Δx equals the reconstruction volume sampling distance. Noting that integrating the footprint over an interval (ti − ti. DCt aliasing during forward projection is avoided by the squared sinc factor as discussed in Section 3.4.j )..10) where ζ(θ) = max(| cos θ|. Thanks to the integration technique presented in (9. However.i ∗ ΠΔt )(ti − ti. we see that the resulting coefficients of the P-matrix are given by pij = Δx (gJ. gJ is integrated along intervals corresponding to the detector sampling distance Δt according to the integration technique described above. t) + ls(θ))dldt = Δ t1 R t2 t1 g(θ.j . (9. i. DCaliasing is avoided both in forward projection and backprojection. and how the ith measurement contribute to the image during backprojection. Λ(t) = (0. t) = 1 Λ ζ(θ) t Δx ζ(θ) . k ∈ Z+ \{0}. we study the coefficients of the P-matrix in more detail. which we refer to as the strip integral method. Furthermore. t1 )). (9. t)dt − t1 −∞ The new projection operator. let ti. t) and ΠΔt is a rectangle function of width Δt .3. instead of sampling gJ as in the original Joseph method. x (9. ti.

15) end end end . t1 − xj ) (see Fig. t2 − xj ) − IP (θ.1: Pseudocode for the strip integral method. The projection of the Joseph footprint onto these lines is always a triangle of length 2Δx . 9.14: Illustration of how the image is divided in to lines {Li } along the coordinate axis that intersect the ray direction with maximal angle. its projection onto the lines {Li } is always a triangle of width 2Δx .3 A strip integral projection method 129       Figure 9. t2 ) Determine which pixels that contribute to the ray for a l l c o n t r i b u t i n g p i x e l s Calculate the contribution from the current voxel as the difference IP (θ. 1 2 3 4 5 6 7 8 9 10 11 12 13 Calculate the integrated footprint table IP (t) for a l l p r o j e c t i o n a n g l e s Divide the image into lines along the coordinate axis maximizing its smallest angle to the ray direction for all image lines f o r a l l image l i n e s Determine the integration interval (t1 .10)).14. Listing 9.1.9. A brief pseudocode description of the algorithm is found in Listing 9. original Joseph method. Because of the stretching of the Joseph footprint (see (9. the first step in the implementation is to divide the image into a set of lines {Li } parallel to the coordinate axis maximally orthogonal to the ray direction as shown in Fig. 9.

B: Transpose PT The distance driven projection method by De Man and Basu [17] is very similar to the strip integral method presented here. 2) P: Joseph.3.13) This is equivalent to the WLS-IFBP method from Chapter 7 with vi = 1. B: Transpose PT 4) P: Distance driven (see below). Four different forward projection/backprojection pairs were compared. namely 1) P: Joseph. 3) P: Strip integral. 2 (9. . Each detector contribution was calculated as a mean value of seven parallel sub-rays. t1 − xj ). Noise-free input data were generated from a Shepp-Logan phantom with a length of 368mm.130 Data acquisition modeling     Figure 9.15: The contribution from the jth pixel to the indicated ray is given by IP (θ. i = 1.. and with optimal choices of the step lengths αk as described in (7. B: Standard backprojector BS .. 2-D reconstructions of the SheppLogan phantom [89] have been studied. Other parameters are listed in Table 9. B: Transpose PT . The method used for reconstruction was a steepest descent method for minimizing the cost function z(f ) = 1 Pf − p 2 N 2 W N +β i=1 j=1 dij (fi − fj )2 . Table 9.2 Experiments For evaluation of the strip integral method. t2 − xj ) − IP (θ. The only difference is that the distance .1. N ..1: Scanning and reconstruction parameters Number of channels Number of projections (1/2 turn) Number of pixels Detector width Width of reconstructed volume 336 & 672 290 & 580 512 × 512 500mm 500mm 9.15).

Since the width of the strip in the strip integral method adapts to the local ray density.3 Discussions and conclusions From the results in the previous section.49mm. and the number of views has been increased from 290 to 580. Although not investigated here. The number of channels is 336 and the number of views per half turn is 290.14) while the strip integral method uses the slightly more complicated triangular footprint gJ given by (9. the image sampling distance is Δx = 500/512mm ≈ 0. the improvement in resolution is expected to be accompanied by increased noise levels as observed for the multiple ray reconstructions in Fig.3. 9. and Δx /Δt = 1.16 shows reconstructed images of the Shepp-Logan after ten iterations. the Joseph method combined with the standard backprojector gives the lowest σe1 value. amplification of noise can probably be avoided without loosing the improved high contrast resolution. indicate that the images in Fig. Δx /Δt ≈ 0. the strip integral method can improve spatial resolution by taking into account the blurring during acquisition in a better way than the original Joseph method. 9.66.16 a) and b). Instead. 9. Therefore. no improvement of spatial resolution can be observed for the distance driven method or the strip integral method. 131 (9. reflecting errors around edges. So far. t) = 1 Π ζ(θ) t Δx ζ(θ) . the σe1 measurements. measured by σe2 is obtained with the strip integral method. Fig.16 c) and d) are sharper than those in Fig.11. i. 9. In Fig.31. 9. Δt equals 0. it is clear that the strip integral method can improve reconstruction results by reducing artifacts due to DC-aliasing in backprojection.3 A strip integral projection method driven projection operator uses the rectangular footprint gDD (θ.2) over (i) the complete image and (ii) regions containing only low contrast structures.75mm. In comparison. 9. or aliasing of false high frequencies in the basis function during forward projection. The lowest artifact level. it is assumed to be especially useful for avoiding aliasing associated with divergent ray geometries such as the semi-parallel geometry or . Furthermore.10). This implies a detector sampling distance Δt = 500/336mm ≈ 1. the σe2 value is lower for the strip integral method than for the other methods.98mm. Furthermore. By combining the strip integral method with the nonlinear regularization presented in Chapter 8.e. However. the number of channels has been increased from 336 to 672.9.17. in contrast to Fig. Also in this case. the strip integral method has only been evaluated on a parallel-beam geometry. 9.16. Image quality was studied by visual inspection and measurement of σe values (see Section 4.

σe1 = 76HU B : PT σe2 = 1. Nch = 336 and Nproj = 290.6HU d) P : Strip int. σe1 measures errors over the whole image. dr.132 Data acquisition modeling a) P : Joseph σe1 = 79HU B : BS σe2 = 1. 40HU). .2HU Figure 9. σe1 = 76HU B : PT σe2 = 1. Greyscale window (0HU.5HU c) P : Dist.4HU b) P : Joseph σe1 = 78HU B : PT σe2 = 1.16: Reconstructed images of the Shepp-Logan phantom after ten iterations. while σe2 measures errors on low contrast regions.

σe1 measures errors over the whole image.17: Reconstructed images of the Shepp-Logan phantom after ten iterations.7HU Figure 9. dr. Greyscale window (0HU. σe1 = 56HU B : PT σe2 = 1.8HU d) P : Strip int. σe1 = 56HU B : PT σe2 = 4. while σe2 measures errors on low contrast regions.3 A strip integral projection method 133 a) P : Joseph σe1 = 55HU B : BS σe2 = 3.3HU c) P : Dist. .9.8HU b) P : Joseph σe1 = 56HU B : PT σe2 = 3. 40HU). Nch = 672 and Nproj = 580.

since this seems inconsistent with the results presented in the original publication [17].17.134 Data acquisition modeling the cone-beam geometry (see Fig. One interesting property of the integration technique (9. (9. and is another possible subject for future research. For example. Then. 125]. but faster and simpler than the more accurate and more aliasing suppressing blob based methods by Mueller et al. More remarkable is the result in Fig. but the fact that the method can be expressed as a series of resampling operations accessing data in a local and predictive way seems to make it a good candidate for efficient implementation on parallel processing hardware such as modern graphics board. 3. for any choice of parameters Δ1 and Δ2 the function (f ∗ ΠΔ1 ∗ ΠΔ2 ) can be evaluated by looking up four values in the f2 table. future research should include investigations on divergent beam geometries. This is explained by the fact that the Fourier transform of gJ contains an extra sinc factor suppressing frequencies above the Nyquist frequency of the reconstructed images. Therefore. Our experiments show that aliasing artifacts caused by the discontinuities in the rectangular footprint gDD of the distance driven method can be reduced by switching to the triangular footprint gJ .15) where f1 (t) = −∞ f (t )dt and f2 (t) = −∞ f1 (t )dt . Efficient hardware implementations of the strip integral method have not yet been considered. 9. assume that we want to calculate the convolution of some function f with the two rectangles ΠΔ1 and ΠΔ2 . we expect the strip integral method to be slower than the distance driven method. having pre-calculated a table of f2 . This could enable efficient implementation of advanced analytical forward projection methods such as those presented by Danielsson and Sunneg˚ ardh in [15].9) is that it can be repeated to efficiently calculate convolutions with several rectangle functions. [124. t t . From a performance point of view. showing stronger aliasing artifacts for the distance driven projector than the Joseph projector.6). (f ∗ ΠΔ1 ∗ ΠΔ2 )(t) = ∫ ∫ t+Δ2 ∫ t +Δ1 f (t )dt dt = = t−Δ2 t −Δ1 t+Δ2 t−Δ2 f1 (t + Δ1 ) − f1 (t − Δ1 )dt = = f2 (t + Δ2 + Δ1 ) − f2 (t + Δ2 − Δ1 ) − f2 (t − Δ2 + Δ1 ) + f2 (t − Δ2 − Δ1 ). One possible explanation could be that the original publication consider a cone-beam geometry where the adaptive anti-aliasing properties of the distance driven projection method should help improving the reconstruction. [68] and Ziegler et al. Thus.

The main drawbacks of these methods are that they require relatively large amounts of memory for look-up tables. [99]. To reduce the number of unique coefficients to be stored. backprojection to a polar grid has previously been suggested [34].1) The corresponding number for the standard backprojector BS is approximately Nx × Ny × Nz × Nproj−irrad ≈ 512 × 512 × 250 × 2300 ≈ 1. Furthermore.1 × 1010 . who suggested backprojection to a stack of axial images rotated with a z-dependent angle φ = φ0 + 2πz/P . since all coefficients are pre-calculated. If the number of angles in the polar grid equals the number of projection angles. (10. A similar trick was suggested by Steckmann et al. For instance. These techniques are very efficient.1 Introduction The computationally most expensive operations in iterative reconstruction are the forward projections and the backprojections.Chapter 10 Implementation on Graphics Processing Units (GPU) 10. the same interpolation and weighting coefficients can be used for all axial images. the number of computations will be huge. the number of non-zero entries in the P and B matrices is O(N 4 ). Given that input and output data dimensions are proportional to some number N . they also allow for complex models of the acquisition process without disproportionate performance penalties. the backprojected images must be . the number of bilinear interpolations needed in the Joseph forward projector P and backprojector PT in the physical head phantom experiments presented earlier in the thesis equals Nch × Nrows × Nproj−tot × Nx = 736 × 32 × 4267 × 512 ≈ 5.2) where Nproj−irrad is the approximate number of projection angles for which a voxel is irradiated.5 × 1011 . even for small problems. Clearly. (10.

and need to be calculated on the fly.136 Implementation on Graphics Processing Units (GPU) resampled to a Cartesian grid. 86]. CUDA is a set of tools for general purpose programming of GPUs. Compared to other well-known libraries such as OpenGL and DirectX. we first take a 1 Our CPU backprojection implementation uses one processor core. especially in the case of iterative reconstruction. 79]. but have the drawbacks of being expensive. . in some cases. and therefore not suitable for research. Therefore. For modern GPUs. 53. To handle the computational load. we present and discuss our experiences from implementing the operators P. and look-up tables must be regenerated for any new Region Of Interest (ROI). Therefore. 3. In the following. Note that the implementations presented here are by no means optimal. 116. The purpose of this chapter is to discuss the possibilities of accelerating different operations with CUDA. 10. we have chosen to base our implementations of forward projection and backprojection on CUDA. the coefficients of the B and P matrices cannot be stored. These implementations are very fast. for experiments on realistic input data. and PT . For example. the language of CUDA is more similar to C. Instead. Without creating and exploiting symmetries and partially abandoning the Cartesian grid as described above. most researchers are referred to perform experiments on desktop workstations or. inflexible. Running experiments on full size clinical data sets on desktop workstations can be painfully slow. In 2006. A more carefully designed single core implementation could be expected to run up to four times faster. Several researchers have reported on successful attempts on reducing the computation times by making use of new parallel processing architectures such as the Cell Broadband Engine (CBE) [43. 87] and Graphics Processing Units (GPUs) [67. one single WFBP backprojection of the physical head phantom data set takes almost four hours on a desktop workstation with one Intel Xeon E5440 processor1 . the company NVIDIA released the Compute Unified Device Architecture (CUDA). rather than to provide optimal implementations. and does not utilize the vectorization capabilities of the processor. supercomputing clusters. results indicate that a one can obtain speedups of between one and two orders of magnitude compared to single core CPU implementations. commercial manufacturers of CT systems use backprojectors implemented either on Application Specific Integrated Circuits (ASIC) or on Field-Programmable Gate Arrays (FPGA) [19.2 The GPU from a CUDA perspective To understand how a GPU can improve the performance of some computationally intensive programs by factors ranging from ten to one hundred. No serious attempts have been made to optimize the implementation. it is important to find simple and inexpensive solutions for reducing the computation time. and requires less prior knowledge about graphics programming. BS .

each SM is equipped with a constant cache. while the CPU uses a larger area for cache and control. In order to fully utilize the hardware.2 The GPU from a CUDA perspective 137   Figure 10.2. it is therefore important to try to avoid programs that often require the parallel processes (threads) to take different paths.1: Usage of chip area on the CPU and GPU. Within each SM. makes the GPU well suited to perform simple but arithmetically intensive parallel computations on large amounts data.1 schematically illustrates how the processor chip area is used on a CPU and a GPU respectively. look at the design of the GPU. and 3-D structures by prefetching data in the close vicinity of data . and compare it with the CPU. 10.10. the ALUs are grouped into several Streaming Multiprocessors (SMs). This property together with the large number of ALUs. However. but on different data. to be further discussed in the next paragraph. Fig. Much of the material in this and the following section has been taken from the NVIDIA CUDA programming guide [75]. 10. which is almost a factor ten higher than the corresponding number for a modern CPU.2. the maximum possible number of Floating Point Operations per Second (FLOP/s) for the NVIDIA GT200 processor is close to one TFLOP/s. Maybe the most noticeable difference is that the GPU devotes more chip area for Arithmetic Logic Units (ALUs). the ALUs work in a Single Instruction Multiple Data (SIMD) fashion. The constant cache speeds up memory access from a 64kB read-only region of the global GPU memory. All SPs run the same set of instructions. The maximum rate at which data that can be read by the GPU from the global GPU memory can be as high as 159Gb/s on modern GPUs. 2-D. 10. The texture cache improves memory access from 1-D. there is a significant latency for random fetches. Therefore. and the so called shared memory as shown in Fig. Redrawn from the CUDA programming guide [75]. As indicated in the figure. a texture cache. it can take 400-600 clock cycles to obtain a requested data value from the global GPU memory. To avoid performance degradation due to this latency. As an example. Each SM has eight scalar processors (SPs) controlled by one instruction unit as shown in Fig.

and calls programs to be executed on the device. A program that is executed on the device is called a kernel. The host executes the program responsible for transferring data to and from the device.1 Programming CUDA Kernels and threads In the context of CUDA.1 shows a simple example of a kernel and how it is called from the host.e. and to be stored on the device memory. To accelerate memory access with shared memory. 10.2: Hardware for CUDA. Redrawn from the CUDA programming guide [75]. accessible for all SPs within the SM for both reading and writing. i. The first five lines defines a kernel that performs a pointwise addition of the elements in the arrays A and B. The third possibility of speeding up memory accesses involve the shared memory.. and will not be discussed in the following sections. Listing 10. it is rather complicated to use. When a call to a kernel is executed. Another interesting property of the texturing hardware is that it also implements 1-D. the global memory . and simple and fast handling of boundary conditions. All arrays are assumed to have the length N. or 3-D linear interpolation. being read.3 10. Each such instance is called a thread. Although this memory can be very useful for some applications. which is a 16kb memory per SM. and stores the result in C.3. several instances of the kernel are started simultaneously on the device.138  Implementation on Graphics Processing Units (GPU) $ # # # "    !  Figure 10. 2-D. the CPU and the GPU are commonly referred to as the host and the device respectively. the programmer has to coordinate manually the loading of data needed by the SPs into the shared memory.

// Kernel invocation vecAddKernel<<<1. Each time a kernel is invoked.3 shows the basic organization of threads.g. the threads have to be organized into several so called thread blocks. Only one grid. For example. and callable from the host only. to specify the dimensions of Grid 1 in Fig. 3. . As the kernel is invoked. 2.3 Programming CUDA 139 Listing 10. } int main() { // Missing: code that sets up the device. i.3.3.e. Fig. and to allow for vectors larger than 512 elements. much like an index variable in a for-loop. The dimensions of the grid and thread blocks are specified in the call to the kernel by passing variables. gridDim and blockDim.B) from // the host to the device. 10. N>>>(A. The built in variable threadIdx on line three tells each thread on which data to work. 10. e.2 shows complete vector addition C/CUDA code that makes use of all available SMs for large input vectors2 . blockDim>>>. 2 The number of elements in the vector must be greater than or equal to the thread block size times the number of SMs on the device of interest. float *C) { int i=threadIdx. To define the kernel.C). one kernel. float *B. The code is a slightly modified version of CUDA programming guide [75]. allocates // memory on the GPU and copies input data (A. N threads are created and executed on one single SM. C[i]=A[i]+B[i]. When the kernel is invoked on line 14.10. This is accomplished by dividing the total number of threads into N/BLOCKSIZE blocks of size BLOCKSIZE (lines 26 and 27). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 __global__ void vecAddKernel(float *A. the function qualifier global is used. } of the GPU. and are executed when the SMs become available.2 Thread organization To make use of several SMs. 1) and blockDim should be set to (4. a new grid of thread blocks is instantiated. gridDim should be set to (3. Listing 10.B. within the triple inequality signs as <<<gridDim. 10. // Missing: code that copies the output data (C) from // the device to the host. each of these blocks executes in parallel if there are enough SMs. can be run at a time on the device.x.1: C/CUDA code for adding two vectors. This means that the following function should be run on the device. The remaining blocks are put in a queue. 1).

Bd. N*sizeof(float). In contrast to Listing 10. float * Cd. Cd. 1). cudaFree(Bd). cudaMalloc((void **) &Bd. 1). cudaMalloc((void **) &Ad. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 #define BLOCKSIZE 256 __global__ void vecAddKernel(float *Ad. 1. float *Ch. Bh. cudaMemcpy(Bd. if (i < N) Cd[i]=Ad[i]+Bd[i].140 Implementation on Graphics Processing Units (GPU) Listing 10. } void vecAdd(float *Ah. all available SMs are used. 1. int N) { int i= (blockIdx. cudaMemcpyHostToDevice). // Allocate and copy memory from host to device float * Ad. cudaMemcpyHostToDevice). N). // Copy the result back to the host cudaMemcpy(Ch. float *Bd. // Set up gridDim and blockDim dim3 blockDim(BLOCKSIZE. // Free device memory cudaFree(Ad). N*sizeof(float). Ah. int N) { // Find and initialize the device with maximum GFLOPS cudaSetDevice( cutGetMaxGflopsDeviceId() ). cudaFree(Cd).x * BLOCKSIZE) + threadIdx. cudaMalloc((void **) &Cd. } . float *Bh. // Kernel invocation vecAddKernel<<<gridDim. N*sizeof(float). cutilSafeThreadSync(). N*sizeof(float)). Cd.x. N*sizeof(float)). float * Bd. dim3 gridDim(N/BLOCKSIZE + (N\%BLOCKSIZE>0 ? 1:0). blockDim>>>(Ad. float *Cd.1.2: Complete C/CUDA code for a function performing pointwise addition of two float arrays. cudaMemcpyDeviceToHost). N*sizeof(float)). cudaMemcpy(Ad.

The . the total execution time was 15. the solution in Listing 10.3 Programming CUDA 141  Figure 10. Therefore.2 may have a negative effect on performance. However. these negative effects on performance are expected to be small.10. For thread blocks containing a mixture of threads that do and do not satisfy the condition. suggesting that the cost of the if-statement is much smaller than the costs of creating threads. and compute the result. sufficiently many blocks are generated to generate an output data with a length that is equal to the input data length rounded up to the nearest multiple of the thread block size. and organization of threads into blocks and grids. Redrawn from the CUDA programming guide (2007) [74]. calculating the index variable. we iterated the kernel invocation 10000 times with N equal to 256 × 32768. To handle vectors with lengths that are not multiples of the thread block size. in most cases the number of threads not satisfying the condition are relatively few. .40s.if-statement on line 6 makes sure that no threads read or write out of bounds.3: Execution of CUDA kernels. Both with and without the conditional statement. To investigate the effect of this conditional statement on blocks for which all threads satisfy the condition.

thus making it possible to use more threads simultaneously on one SM. the device has a total constant memory of 64kB. In CUDA. textures are optimized for 2-D locality. as well as the operations discussed in the following sections. or by extrapolation with repeated copies of the texture. neighborhood operations. this memory can be used for certain relatively small precalculated lookup tables. and provides simple handling of boundary conditions as well as linear interpolation. or tex3D). According to the CUDA programming guide. meaning that improved performance is obtained if threads in the same block or neighboring blocks read from neighboring locations in the texture.4 Constant memory As mentioned in Section 10. but it cannot be directly read from or written to. 10. a texture is a chunk of memory that is connected to the texture hardware via a texture reference. When a cudaArray is created in device memory and a texture reference is bound to it. Texture references can be bound to 1-D. This excessive amount of simultaneous threads per SM is used by CUDA to hide memory access latencies.2. The device has a mechanism for rapidly switching between one set of threads and another. Threads that are not active have all data stored in registers. A cudaArray can be accessed by copy commands and texture fetch commands. access to the constant memory is as fast as access to a register. which is much larger than the number of SPs per SM.142 Implementation on Graphics Processing Units (GPU) In Listing 10. The interpolation can be set to nearest neighbor or linear interpolation. More specifically. On a cache hit. Therefore. Each SM has a cache of 8kB.3. this condition is easy to meet. To get most out of the texture functionality. taking the texture reference and a coordinate vector. One way to improve performance of a CUDA program is therefore to minimize the number of required registers per thread. 10. the type and dimensionality of data are specified. which amount to 8192 per SM on the G80 chips and 16384 on GT200 chips. as well as the interpolation to use and how boundaries should be handled. tex2D. The texture hardware improves the performance of read access by caching. 2-D. a value from the texture is obtained by using a texture fetch function (tex1D. In resampling operations. the number of threads per block must be smaller than the total number of registers on the SM divided by the number of registers used by one thread. the thread block size equals 256. and the boundaries can be handled either by constantly extrapolating the boundary pixels of the texture. Therefore. each SM provide a so called constant cache. the memory connected to the texture should be stored in a special structure called cudaArray.3 Textures Since textures are an important part of the implementations in the following sections. we briefly explain them here. .3. In the kernel.2. and 3-D arrays.

This allows us to use the caching mechanisms of the texture hardware. end The first step in the host code is to allocate the required memory on the host and the device. the corresponding source position has to be calculated.4 Forward projection Implementation of the Joseph forward projection in CUDA is straightforward. The threads are grouped into blocks of 16 detector channels times 2 detector rows.z).e. Given parallel projection angle θ. i. and bound to a texture reference. using a 3D texture makes the implementation simpler by avoiding use of several texture references. for a l l p r o j e c t i o n a n g l e s Calculate source position for the central ray (αC . The last preparational step of the host program is to calculate lookup tables that simplify calculations in the kernel.4 Forward projection 143 10.3: Pseudocode for the forward projection. On the device. 1 2 3 4 5 6 7 8 9 Copy the input volume to the device.3) where t0 is the t-value corresponding to chIx=0. We begin explaining the program to be run on the host.3. as well as the hardware linear interpolation. To reduce the number of registers required by the kernel. and Δt is the sampling distance along the t direction. We let one thread correspond to one output value.2.3. For simplicity. it is assumed that there is enough space on the host and device to store the whole voxel volume. and that only one line integral per measurement is calculated. However. Pseudocode for this program is listed in Listing 10. and one grid represents one projection angle. Host part. 2. Call the kernel which computes a projection image for the current angle. zC ). it is not necessary to use a 3D texture. and by avoiding the division of the implementation into two cases depending on the direction of interpolation. Since only 2D interpolation is required by the Joseph method. Configure grid and thread blocks. (10. one ray. Copy the projection image from the device and save it.10. the rightmost term can be pre-calculated and made accessible to the kernel through . For each ray/thread. Calculate constant memory lookup tables for simplified computation of the x-ray source position (α. the input volume is stored as a three-dimensional cudaArray. the α-angle (see Fig. The second step is to set the gridDim and blockDim parameters that specify the thread organization described in Section 10. Listing 10. and bind a 3D-texture to it.6) of a ray with channel index chIx can be calculated as α = θ − arcsin(t/RS ) = θ − arcsin((t0 + chIx × Δt )/RS ). and to load the input volume.

Calculate a normalized ray direction vector Transform the source position and normalized ray direction to the coordinate system of the voxel volume. Pseudocode for the kernel is listed in Listing 10. The computation times for the clock phantom were 5800s for the CPU implementation and 81s for the GPU implementation. In other words. As previously mentioned. Save measurement to the device memory. This code correspond to the calculation of one line integral. In the kernel invocation. a lookup table for z = zC + (α − θ)P/(2π) can be pre-calculated. the task of one instance of this program is to calculate one line integral. Accumulate the contribution. Calculate source position with help from the pre-calculated constant memory tables. Device/kernel part.4: Pseudocode for the forward projection. except from the interpolation step on line 10. f o r a l l c o n t r i b u t i n g image p l a n e s Calculate intersection between the ray and the contributing plane. the result is copied to the host memory and saved to file. Listing 10. When all preparations are finished.4. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Calculate channel and row indices (chIx. the main loop over all projection angles starts.144 Implementation on Graphics Processing Units (GPU) the constant memory. In CUDA. To examine the performance of the implementation. Calculate the first and last intersection between the ray and the contributing image planes. where P is the patient table feed per turn. which performs the actual interpolation. this reduces to a call to the texture hardware. The line of interest is specified by the built in variables blockIdx and threadIdx variables. Fetch a contribution from the 3D texture bound to the input volume. The rest of the kernel program does not differ much from the inner loop in a normal CPU implementation.rowIx) from blockIdx and threadIdx. In the same way. these variables are sent as parameters to tell the kernel which projection image that shall be calculated. end Perform length scaling. the θ angle and the z for the x-ray source corresponding to the central ray are calculated. the pitch of the helix. performance improvements of a factor of 72 and a factor of 84 were observed. or equivalently. . As the calculation is finished. and an NVIDIA Tesla C870 computing card. Corresponding numbers for the head phantom were 2700s and 32s respectively. we generated projection data from the clock phantom and the head phantom from previous chapters. For each angle. on a system with an Intel Xeon E5440 processor. which in an initial step are transformed into a channel index chIx and a row index rowIx.

2 (θ) Tch. and q. Δt Δt Δt Tch.5. the CUDA implementation of the WFBP and standard backprojector is similar to the CUDA implementation of the Joseph forward projector.5 10. and calculating lookup tables. The detector row index rowIx is calculated as rowIx = q0 RS (z − zS ) − = Δq m Δq 1 = z0 + zIx × Δz − zS0 − Δ l Δq cos β − Rq S Trow.1 .3 (θ) (10. Tch.2 (t) − q0 .1 Backprojection The WFBP/standard backprojector In many ways. Copy the projection image from the device and save it. Host part. one thread block corresponds to 8 × 8 neighboring threads in an axial image. Configure grid and thread blocks. To simplify the calculation of the projection of a voxel onto the virtual detector. Δq (10. 10.4). Equivalent relationships apply to y. Pseudocode for the host part is listed in Listing 10.1 (t) βP 2π Trow. f o r a l l a x i a l images Calculate the z-coordinate of the image. the calculation of the channel index chIx is done according to chIx = Δx −Δx x0 sin θ − y0 cos θ − t0 +xIx sin θ +yIx cos θ. t.5) . Fig. Calculate constant memory and texture lookup tables for simplified computations of voxel projections onto the virtual detector (Fig.3 indicate pre-calculated constant memory tables that depend on the variable θ. In our WFBP/standard backprojection implementation.5 Backprojection 145 Listing 10.2 .5: Pseudocode for the WFBP/standard backprojector.10. and bind a 3D-texture to it. the preparations consist of loading input data to the device and bind it to a 3D texture. both constant memory lookup tables and texture lookup tables are used. 10.5.blockDim). and Tch. and one grid consists of all threads needed to create one axial image. set up the parallelization (gridDim.1 (θ) Tch. As for the forward projection implementation. end 10.4) where Tch.4 shows the backprojection geometry. Note that x and xIx are related by the equation x = x0 + xIx × Δx . one thread corresponds to one voxel. In the kernel. Call the kernel which computes the axial image. 1 2 3 4 5 6 7 8 9 Copy the input projection data to the device.

.2 are t-dependent tables that are stored in the texture memΔ l ory.3 . the voxel is projected onto the virtual detector. The term Rq is given by S Δq l Δq Δ q Δx Δ q Δx = (x0 cos θ + y0 sin θ) +xIx cos θ +yIx sin θ .4 . the inner loop covers ˜ all source positions corresponding to this angle. and the angle θ + π. To examine the performance of the backprojector. The .4: The backprojection geometry. Listing 10. and passed as an argument to the kernel invocation.4 (θ) Trow. When the device code has produced an image. and Trow. it is copied back to the host and saved to file. Trow.5 (θ) (10. For each source position. The main loop reconstructs one 2D-image at a time. The first step in the kernel is to calculate the voxel indices xIx and yIx from blockIdx and threadIdx. the corresponding z-value is calculated. using the scanning geometry described in Chapter 4. Lines marked WFBP perform the weighting and normalization that are specific to the WFBP method. For each image. Trow. the channel index chIx and row index rowIx are calculated. i. the detector value is read and accumulated. If the projection of the voxel lies within the boundaries of the virtual detector.6) Similarly to the channel index calculation.1 and Trow.e.6 shows pseudocode for the kernel part of the WFBP/standard backprojector.3 (θ) Trow.146  Implementation on Graphics Processing Units (GPU)           Figure 10. For each projection angle θ.5 represent θ-dependent constant memory tables. the clock and head phantoms were reconstructed. This is followed by the main loop running over all ˜ ˜ projection angles θ from 0 to π. where Trow. RS RS RS RS Trow.

this row turns into a column in PT . i f rowIx >0 and rowIx<Nrows−1 Fetch a contribution from the 3D texture bound to input data WFBP: apply Q-weight WFBP: accumulate Q-weight end end WFBP: Divide the pixel value with the accumulated Q-weight end computation times were 25000s and 15000s for the CPU implementation.5. correspond to one thread as in the previous section.5 Backprojection 147 Listing 10. This row holds the coefficients describing how the different voxels in the voxel volume contribute to one projection data element during forward projection. PT seems to be more difficult to implement efficiently in CUDA than P and BS . ˜ θ + 360◦ . one voxel. as the voxel volume contribute to this projection data element during forward projection with P. When P is transposed. This column defines how PT maps one projection data element into the voxel volume.2 The Joseph backprojector PT To understand what the Joseph backprojector does. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Calculate xIx and yIx from blockIdx and threadIdx. θ + 900◦ . First. i f rowIx >0 and rowIx<Nrows−1 Fetch a contribution from the 3D texture bound to input data WFBP: apply Q-weight WFBP: accumulate Q-weight end end ˜ Calculate chIx for θ = θ + 180◦ (see (10. ˜ + 180◦ .4)). This corresponds to speedups of 260 and 180 times respectively. In other words.5) and (10. ˜ ˜ f o r θ = θ. . ˜ ˜ for θ = θ Calculate rowIx (see (10.5) and (10. This approach is difficult because the task of finding the rays contributing to the output voxel becomes very .10.6)).6)).. ˜ f o r θ = 0◦ to 180◦ ˜ Calculate chIx for θ = θ (see (10... one projection data element contributes to the voxel volume with the exact same coefficients. consider the possibility to let one output data element. θ + 720◦ . and 96s and 84s for the GPU implementations.. Calculate rowIx (see (10. during backprojection with PT . i. θ + 540◦ .4)). 10. we begin considering a row in the matrix of the Joseph forward projection P. . Device/kernel part.6: Pseudocode for the WFBP/standard backprojector.e.

148 Implementation on Graphics Processing Units (GPU)     Figure 10. one thread corresponds to one voxel update of one ray. complicated even for simple ray geometries. let one thread correspond to one ray. To solve this problem. 10.6: Division of one projection image into blocks to be executed in a sequence. Corresponding times for the head phantom were 3200s for the CPU and 750s for the GPU.5: The difference between Joseph forward projection and backprojection. the rays are sufficiently spatially separated so that no voxel update conflicts occur. thus allowing for parallel execution.2 times. . The second possibility would be to reuse the procedure of the forward projection. Since the number of rays per block is much smaller than the total number of rays. Within each block. but replace the reading commands with writing commands (see Fig. In our implementation.6.e. Unfortunately. the performance of the above described approach is poor.     Figure 10. The computation times for the clock phantom were 7200s for the CPU and 2700s for the GPU.7 times and 4. i. the thread block configuration must be altered. This would be a simple and good approach for a single core CPU implementation. This corresponds to speedups of 2. we divide the rays of one projection angle into blocks to be executed in a sequence as illustrated in Fig. and one grid corresponds to one “sequential” block. 10. one thread block corresponds to the total update of one ray. but does not work for parallel implementations since there would be conflicts between different threads wanting to update the same voxel.5).

speedups of more than a factor of 180 were observed. For the forward projection.6 Discussions and conclusions 149 10. and it has been shown that the performance of GPU implementations may vary much with the configuration of parallelization [84]. Two important factors contributing to this improved performance is (i) the hardware interpolation offered by the texture hardware. In the CUDA implementation of the transpose PT of the Joseph forward projection. and (ii) the caching provided by the texture hardware. Since results from previous chapters have indicated that PT is useful in iterative reconstruction. an interesting future research topic would be to search for more efficient exact or approximate implementations of this operation. The CPU implementations could be improved by using SSE extensions and utilizing more cores. each thread performs uncached read and write operations to four different locations in the device memory with rather poor performance. and for the backprojection. When interpreting the speedup factors presented here. Nevertheless. it is clear that graphics hardware can be used to accelerate time consuming operations in reconstruction considerably.6 Discussions and conclusions The results from the previous sections show that the computational performance of the Joseph forward projection and the WFBP/standard backprojection can be improved significantly by using the GPU. .10. we did not find any good use of the hardware interpolation or the texture caches. Instead. one must keep in mind that both the CPU and GPU implementations are sub-optimal. speedups of around a factor of 80 were observed.

150 Implementation on Graphics Processing Units (GPU) .

and to improve the convergence of the methods. Furthermore. but with linear regularization. these effects can be alleviated or completely avoided.Chapter 11 Conclusions and future research 11. the ordered subsets technique [39] can be applied to RIWFBP. To reduce the number of required iterations. evident from comparing the rate of change. but they can also alter spatial resolution and noise properties. Without regularization. In Chapter 6.e. each iteration increases the amount of high frequencies and noise in the reconstructed images. and the WLS-IFBP method based on gradient descent minimization of a spectrally weighted quadratic cost function. In the course of events we found that even with regularization. measured as log10 ( fk+1 − fk /C) in Fig. the amplitude of windmill artifacts is also reduced.15. linear and non-linear regularization have been investigated. 5. it . we have examined possibilities to further improve image quality by accurate modeling of the projection operator. the RIWFBP reconstruction results suffer from overshoots in the z-direction. the OS-IWFBP. To avoid such unwanted effects. cone artifacts can be perfectly suppressed within three iterations of the RIWFBP method. The main conclusion of Chapter 5 is that for cone angles up to ±2. Another positive effect of the regularization is the improved convergence properties. we have developed and examined three IFBP methods for iterative improvement of image quality in helical cone-beam CT: the RIWFBP method based on iterative application of the WFBP method.1 Summary of conclusions In this thesis. these can be avoided by preprocessing of projection data with a low-pass filter in the q-direction. Fortunately. i.78◦ . All three methods suppress cone artifacts and alleviate windmill artifacts. linear smoothing of the reconstruction result in each iteration. Thanks to better handling of redundant data. resulting in the OS-IWFBP method.

and therefore reduces aliasing artifacts in comparison with line-integration based projector/backprojector pairs. i.152 Conclusions and future research is shown that this technique can improve the rate of cone artifact suppression for high cone angles (4. Due to the non-symmetric QP matrix. we present the WLS-IFBP method. but also more advanced algorithms such as the conjugate gradient method. Therefore.6◦ ). noise with an amplitude that is lower than a certain threshold is suppressed. it is described how to use nonlinear regularization with the RIWFBP and WLS-IFBP methods. and they could possibly be compensated for by a more sophisticated weighting. the RIWFBP method is difficult to formulate in terms of an optimization problem. PT has the advantage of producing exactly the structures P can detect. However. we present a projection operator based on strip integrals. since the strip integral to some degree takes into account the blurring . Thanks to the latter feature. lowpass filtering adapted to the projection data sampling is introduced. and that the true optimum might not be found due to the fixed step length α. convolution with a rectangle adjusted to the ray density (detector sampling). This lowpass filtering helps to avoid DC-aliasing during backprojection with PT . in its current form. Therefore. and hence difficult to analyze. However. the OS-IWFBP is non-regularized. Furthermore. WLS-IFBP can be expressed as a gradient descent method for solving a weighted least squares problem. In the second part of Chapter 9. and its convergence properties are uncertain. Nonlinear regularization can be used to obtain contrast dependent spatial resolution. Thereby not only improvements such as automatic determination of the step lengths αk . Unfortunately. despite this fact. which is similar to the RIWFBP. Furthermore. However. a considerable improvement is observed during the first ten iterations. the difference is very small.e. Because of the nonlinear regularization. suppresses DC-aliasing in both forward projection and backprojection. while noise with an amplitude that is higher than this threshold is amplified. and to compare with other iterative methods for tomographic reconstruction. Experiments with WLS-IFBP show that the spatial resolution for high contrast structures can be improved while preserving the spatial resolution and signal-to-noise ratio for low contrast structures. certain artifacts are slightly more pronounced for WLS-IFBP than for RIWFBP. it is shown that accurate modeling of the acquisition process can improve image reconstruction in terms of spatial resolution. but has a simplified normalization procedure. to improve. One problem with nonlinear regularization is that more iterations are needed for convergence. The main conclusion is that the strip integration. the regularization parameter must be chosen with respect to the expected noise level. backprojection with PT produces better results for more accurate data acquisition models.8◦ and 9. In Chapter 7. In Chapter 8. An interesting observation is that by modeling detector element size and gantry rotation in P. can be introduced. The actual amount of improvement depends on the scanner geometry and on the type of backprojector used. In the first part of Chapter 9.

Hence the gain in performance for implementation of PT stayed at the not overly impressive factor of four. Improved weighting for WLS-IFBP. Unfortunately. the WLS-IFBP generates slightly more pronounced artifacts than the RIWFBP method in some cases. Our experiments on CUDA implementation of these operations in Chapter 10 have shown that the Joseph projector and the standard backprojector can be accelerated with a factor of more than 80 compared to a single core CPU implementation. We mention this particular method because. If this presumption could be proved correct. it may even happen that the optimum cannot be reached by using a constant step length α. Many authors have shown that by taking into account the dependence between variance . It is not obvious that a step length α that is constant for the whole reconstruction volume gives the highest rate of convergence. we have not managed to make use of texture caches or hardware interpolation.2 Suggestions for future research 153 that occurs during data acquisition. Statistical weighting. it has the attractive property of requiring only one calculation of P and PT per full iteration. the WLS-IFBP could be improved by replacing the current Q-weighting with a more advanced weighting that takes data redundancies into account. the transpose PT of the Joseph operator is more complicated to accelerate with CUDA. preliminary experiments have shown that the number of required iterations can be reduced by using higher step lengths close to high contrast structures. the strip-integration projector/backprojector pair also improves spatial resolution. The computationally most expensive part of a reconstruction pipeline is the projection and backprojection operations. Thanks to the line search. As shown in Chapter 7. For instance. However. All methods investigated in this thesis falsely assume that the variance of a detector measurement is independent of its expectation.2 Suggestions for future research Conjugate gradient optimization. Unfortunately. Together with a line search method.11. In the case of nonlinear regularization. One possible cause for this could be that the normalization in the RIWFBP method takes better care of the actual data redundancies. We believe that both the final convergence and the rate of convergence may be improved by a more advanced update method. 11. some of the methods for image quality improvement proposed in this thesis increase the number of required iterations. such a method would also eliminate the need for choosing a suitable α depending on the regularization parameter β. this could improve the convergence rate without risking divergence. such as the Conjugate Gradient (CG) method with line search by Fessler and Booth [25]. similarly to the methods studied in this thesis. Spatially variable step lengths.

there is hope to improve spatial resolution and suppress in-plane aliasing artifacts. As previously mentioned.154 Conclusions and future research and expectation.3. it also has the advantage of producing exactly the structures P can detect. In Chapter 9. in-plane resolution is inevitably lost. 126]. it is of interest to find good projector/backprojector pairs P/PT such that both P and PT can be efficiently implemented on existing hardware. a good approximation of both projector and backprojector can easily be implemented in CUDA. an improved signal to noise ratio can be obtained [109. Unfortunately. Therefore. Although we have not found an efficient exact implementation of this projection/backprojection pair. and thereby entirely avoid rebinning. Since rebinning to semi-parallel data is the first step in the methods studied here. and ensuring that BP is positive definite. . it was shown that using the adjoint matrix B = PT for backprojection sometimes results in better spatial resolution. An interesting topic for future research is to examine computationally inexpensive possibilities to improve the noise models of IFBP methods. Native cone-beam geometry. By formulating the WLS-IFBP method directly in the cone-beam geometry. Implementation friendly projection/backprojection pairs. this comes to the price of a more complicated numerical problem. One example of such a projector/backprojector pair is presented in 9.

p. p. p. p. p.Appendices A Acronyms AMPR ART ASIC ALU CBE CMIV CUDA CT CG CRC DFM EMI FPGA FBP FLOP FWHM GPU HU ICD IFBP LROI MRF MAP ML MLEM p. p. p. p. p. p. p. p. p. 67 21 136 137 136 2 7 1 153 22 11 1 136 3 137 59 7 4 67 4 49 34 22 22 22 Adaptive Multiple Plane Reconstruction Algebraic Reconstruction Technique Application Specific Integrated Circuit Arithmetic Logic Unit Cell Broadband Engine Center for Medical Image Science and Visualization Compute Unified Device Architecture Computed Tomography Conjugate Gradient Contrast Recovery Coefficient Direct Fourier Method Electrical and Musical Industries limited Field-Programmable Gate Array Filtered Backprojection Floating Point Operation Full Width at Half Maximum Graphics Processing Unit Hounsfield Units Iterative Coordinate Descent Iterative Filtered Backprojection Local Region Of Interest Markov Random Field Maximum a Posteriori Maximum Likelihood Maximum Likelihood Expectation Maximization . p. p. p. p. p. p. p. p. p. p. p.

p. y. p. p. p. p. p. 3 23 9 14 15 10 10 2 3 3 36 3 30. β q u. v ρ RS h P βmax κmax Nx .156 Modulation Transfer Function Ordered Subsets Ordered Subsets Convex Ordered Subsets IWFBP Point Spread Function q-Generalized Gaussian Markov Random Field Region Of Interest Regularized Iterative Weighted Filtered Backprojection Root Mean Squared Error Scalar Processor Signal to Noise Ratio Simultaneous Algebraic Reconstruction Technique Simultaneous Iterative Reconstruction Technique Single Photon Emission Computed Tomography Slice Sensitivity Profile Streaming Multiprocessor Weighted Filtered Backprojection Weighted Least Squares IFBP Appendices MTF OS OSC OS-IWFBP PSF qGGMRF ROI RIWFBP RMSE SP SNR SART SIRT SPECT SSP SM WFBP WLS-IFBP p. p. their meaning at a certain location should be clear from the context. p. Scalars: geometry and sampling Reconstruction space coordinates Reconstruction space coordinate vector Parallel-beam projection data coordinates Fan-beam projection data coordinates Projection data row coordinate Frequency coordinates Radial frequency coordinate Source-isocenter distance Detector height Table feed per turn/helical pitch Maximal fan-angle Maximal cone-angle Voxel volume dimensions (Ny = Nx ) Sample distances for voxel volume (Δy = Δx ) x. t α. p. p. p. p. 27 . Unfortunately the symbols α and β mean both fan-beam projection coordinates and parameters in the RIWFBP method. p. Nz Δx . Δz p. p. p. p. p. p. p. p. z r θ. p. p. p. However. p. 135 23. p. p. p. p. 40 6 22 6 11 7 50 45 37 137 20 21 21 22 40 137 4 7 B Notation Below follows a list of symbols and acronyms used in the thesis. p. p.

p. Nrows Δβ . R(·) F Ft F2 R p. y. 38 38 38 57 17 34 48 48 157 p. Ω2 ρ50% Q β βxy βz α. p. Δq Δt . f∞ R. p. p. p. f (r) V (·) Q P pin f1 . θ) b(r) wi (r) f (x. 115 p.B Notation Detector dimensions Sample distances. p. non-rebinned proj. 12. rebinned proj. data Scalars: evaluation Error measurement (RMSE) Noise measurement Set for error and noise measurements MTF 50% frequency Scalars: reconstruction parameters Detector row down-weighting parameter Regularization parameter Regularization parameter xy-plane Regularization parameter z-direction Scalars: others Step lengths Scalar valued functions Projection data (2D) Basis function Irradiation function Function to be reconstructed Potential function Vectors and matrices Reconstruction matrix Projection matrix Input projection data vector Sequence of reconstruction results Limit point for reconstruction results Regularization matrix/operator Transforms 1D Fourier transform 1D Fourier transform (radial direction) 2D Fourier transform Radon transform Nch .. f2 . p. p. αk p(t. z). p. p. . data Sample distances. Δq σe σn Ω1 . p. 9 23 24 6 34 6 6 6 5 48 47 10 10 10 9 . p. p. 36 p.. 19 p. p. 115 p. p. p. 6. p. p. p. p. p.

158 Appendices .

2008. H. Beekman and C. J. The Fourier transform and its applications. Chen. Chang. Herman. Basu. McGraw-Hill. 2008. [9] G.-H. Kak. Edholm. T. Riddle. R. Benson. A method for attenuation correction in radionuclide computed tomography. In 4th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. Flohr. 3D filtered backprojection for curved detector axial cone-beam data on a Playstation 3. 46(7):1835–1844. 35(2):660– 663. 1978. [5] R. 1984. and M. M. Pennsylvania.C. In IEEE Nuclear Science Symposium Conference Record. Raupach. 2001. 1967. . [6] R. [8] L. Danielsson. 1985. and S. Physics in Medicine and Biology. IEEE Transactions on Nuclear Science. Andersen and A. P. Elfving. Pittsburgh.Bibliography [1] A. USA. Dixon. Tang. [2] F. 27(12):1685–1703. Inversion of fan-beam scans in radio astronomy. Medical Physics. [10] P. pages 3522–3525. 1986. 1997. Magnusson. W. 7(1):108–123. [3] T. C. Simultaneous algebraic reconstruction technique (SART): a superior implementation of the ART algorithm. Prior image constrained compressed sensing (PICCS): a method to accurately reconstruct dynamic CT images from highly undersampled projection data sets. [7] Y.N. A method of iterative data refinement and its applications. and J. [4] A. Astrophysical Journal. and G. Singapore. Hornegger. Towards exact 3D-reconstruction for helical cone-beam scanning of long objects. T. Kamphuis. 2007. N. Leng. J. Wavelet Based Noise Reduction in CT-Images Using Correlation Analysis. Censor. A new detector arrangement and a new completeness condition. Eriksson. T. 25(1):638–642. Bracewell and A.-E. 6(1):81–94. 150:427–434. Mathematical Methods in the Applied Sciences. IEEE Transactions on Medical Imaging. Borsdorf. Ultrasonic Imaging. Ordered subset reconstruction for x-ray CT. and S. J. volume 5. Bracewell.

Physics in Medicine and Biology. Nuyts. P. Link¨ping University. De Man. A proposal for combining FBP and ART in CT-reconstruction. In IEEE Nuclear Science Symposium Conference Record. In 8th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. S. Utah. J. G. pages 2708–2710. De Man and S. Marchal. Dupont. Suetens. The 2D-case. P. [15] P. Nuyts. ISSN 1400-3902. volume 5. Iterative reconstruction for reduction of metal artifacts in computed tomography. Lindau. [16] B. . and P.160 Bibliography [11] P. Basis and window functions in CT. Reduction of metal streak artifacts in X-ray computed tomography using a transmission maximum a posteriori algorithm. IEEE Transactions on Nuclear Science. In 7th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. Salt Lake City. IEEE Transactions on Medical Imaging. [17] B. H. De Man. In 9th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. J. Technical Report LiTH-ISY-R-2085. Sunneg˚ ardh. Basu. ISSN 1400-3902. J. and H. Magnusson. 5(5):740–753. C. 20(10):999–1008.-E. Advanced linear modeling and interpolation in CT-reconstruction. De Man. High performance image reconstruction and implementation. J. 2007. 2001. Danielsson and M. Danielsson. Fessler. [22] A. Department of Electrical Engineering. Danielsson and J. Magnusson. Technical Report LiTH-ISY-R-2634. Analysis of an iteration scheme. Suetens. Basu. [18] B. [13] P. 2003. Edholm.A. [19] B.-B.-E. Phd thesis.-E. and P. Dupont. 1996. M. A fast and accurate Fourier algorithm for iterative parallel-beam tomography. Combining Fourier and iterative methods in computer tomography. Link¨ping University. Danielsson. Bresler. 2004. Danielsson and M. The original PI-method for helical cone-beam CT. J. Sunneg˚ ardh. 2001. o [14] P. 47(3):977–981. An iterative maximum-likelihood polychromatic algorithm for CT. and J. In 9th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. G. Thibault. December 1998. J.-E. Germany. A study of four minimization approaches for iterative reconstruction in X-ray CT. Marchal. Hsieh. 2005. Magnusson. Drapkin. Turbell. IEEE Transactions on Image Processing. Department of Electrical Engineering. and K.-E. M. R. France. Sauer. Eriksson. 2000. Katholieke Universiteit Leuven. Lindau. Bouman. o [12] P. Germany. Nilsen. USA. De Man. and E. Distance-driven projection and backprojection in three dimensions. [21] B. 49(11):2463–2475. De Man. March 2005. P. 2007. Magnusson. [20] B. Delaney and Y. 2005. St Malo.

H. Bruder. Hein. Elbakri and J. 21(2):89–99. Erdo˘an and J. A. R. Image reconstruction and image quality evaluation for a 64slice CT scanner with z-flying focal spot.W Wecksung.de/phantoms/index. Stierstorfer. Flohr. A.html. Society for Industrial and Applied Mathematics. 1985. Fessler and S. Fessler. [36] G. 1972. 2002. Algebraic reconstruction techniques (ART) for three-dimensional electron microscopy and x-ray photography. 1988. T. 2005. Bender. The fundamentals of computerized tomography. E. New York. Booth. As available electronically in May 2009 at http://www. [34] T. 1999. Applied Optics. [25] J. Journal of Theoretical Biology. 24(23):4028–4039. Hanson and G. Ordered subsets algorithms for transmission g tomography. Gordon. Simulation of noise. 29(3):471–481. Simon. [35] I. Taguchi. IEEE Transactions on Image Processing. T. [31] R. . Statistical methods for tomographic image reconstruction. K. A. Academic Press. 1970. 35(1):615–619. R.Bibliography 161 [23] I. [28] T. and Silver M. 1987.-C. N. Bruder. Fessler.uni-erlangen. Herman. Flohr. Image reconstruction and image quality evaluation for a 16-slice CT scanner. H. Stierstorfer. US Patent 2006/0029285. Gilbert. and G. Singh. Bulletin of the International Statistical Institute. 44(11):2835–2851. 2003. A. 1980. [32] P. Local basis-function approach to computed tomography. Conjugate-gradient preconditioning methods for shift-variant PET image reconstruction. D. 1999. Primak. 36(1):105–117. Medical Physics. [26] T. A. February 2006. [27] T. McCollough. Polacin. Journal of Theoretical Biology. S. Method for helical windmill artifact reduction with noise restoration for helical multislice CT. and S. Rank-deficient and discrete ill-posed problems: numerical aspects of linear inversion. A. Image reconstruction from projections. Ulzheimer. McClure. and M. A. D. Iterative methods for the three-dimensional reconstruction of an object from projections. Fast MLE for SPECT using an intermediate polar representation and a stopping criterion. [30] P.M. Geman and D. Herman. T. 1997. Statistical image reconstruction for polyenergetic X-ray computed tomography. Fuchs. K. [29] S. H. [24] H. 52:5–21. Hansen. Hebert. Leahy.imp. IEEE Transactions on Medical Imaging. Medical Physics. [33] K. and C. Physics in Medicine and Biology. J. K. 8(5):688–699. 30(5):832–845. 32(8):2536–2547. IEEE Transactions on Nuclear Science. Schaller.

Shaughnessy. Tang. 2002.S. 2005. New York. image quality. Kamphuis and F. 2003. R. 22(5):569– 579. Turbell. volume 6. Computed tomography. 27(4):754–772. and M. 3(4):233–236. [44] M. 1(3):192–196. [48] A. 17(6):1101–1105. 1982. Hyperfast Perspective ConeBeam Backprojection. Advanced single-slice rebinning in cone-beam spiral CT. Bockenbach. Nilsen. [47] C. Beekman. [45] A. IEEE Press. pages 1679–1683. Grass. M. IEEE Transactions on Medical Imaging. pages 2739–2742. 2 edition. IEEE Transactions on Medical Imaging. Nielsen. 2003. H. [41] P. F. USA. In IEEE Nuclear Science Symposium Conference Record. and O. Proksa. Kachelriess. volume 3. Washington. 2003.M. C. October 2006. Schaller. 62(6):2012– 2026. Slaney. Accelerated image reconstruction using ordered subsets of projection data. K¨hler. SPIE Press. [46] W. design. and T. 2005. Conjugate cone beam backprojection for high z-resolution reconstruction. A projection access scheme for iterative reconstruction based on o the golden section. C. Larkin. Utah. 1976. applications. 1998. Salt Lake City. In IEEE Nuclear Science Symposium Conference Record. Hsieh. Medical Physics. [43] M. In 8th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine.162 Bibliography [37] J. Judy. [40] M. Wang. [49] T. IEEE Transactions on Medical Imaging. Kalender. [39] H.-B. [38] J. . Theoretically exact filtered backprojection-type inversion algorithm for spiral CT. X. artifacts and recent advances. Principles of computerized tomographic imaging. In IEEE Nuclear Science Symposium Conference Record. 2004. S. 2000. In IEEE Nuclear Science Symposium Conference Record. Jiang and G. Kak and M. J. Principles. Accelerated iterative transmission CT reconstruction using an ordered subsets convex algorithm. 13(4):601–609. SIAM Journal on Applied Mathematics. Knaup. [50] T. Joseph. Computed Tomography. 2000. pages 3961–3965. J. and W. K¨hler. Hudson and R. volume 4. Convergence studies on iterative algorithms for image reconstruction. K¨hler. pages 15/113–15/115. Katsevich. An improved algorithm for reprojecting rays through pixel images. Medical Physics. and Roy A. M. Erlangen. Thibault. [42] P. The line spread function and modulation transfer function of a computed tomographic scanner. volume 2. IEEE Transactions on Medical Imaging. SNR-weighted ART applied to o transmission tomography. Kachelriess. Kalender. system technology. A. Fundamentals. 1994. A. [51] T. Publicis. 1988. Efficient forward projection through o discrete data sets using tri-linear interpolation. Hsieh.

[56] K. J. Bruder. J. 7(10):1834–1846. 2007. [61] R. [53] J. Landweber. Linogram and other direct Fourier methods for tomographic reconstruction. K. IEEE Transactions on Medical Imaging. Journal of the Optical Society of America A. Multidimensional digital image representations using generalized Kaiser-Bessel window functions. Sweden. Lewitt. M. Pre-processing of projections a for iterative reconstruction. Ruth.imp. 8(2):306–316. [55] L. 2005. Crawford. B. USA.-E. Handling of long objects in iterative improvement of nonexact reconstruction in helical conebeam CT. Beekman. Alternatives to voxels for image representation in iterative reconstruction algorithms. 1990. Z. London. [63] S. [58] G. Magnusson. Lange and J.uni-erlangen. G. Carson. 25(7):935–940. [57] K. Hsieh. Li. S. P. 73(3):615–624. [64] M. Physics in Medicine and Biology. Evaluation of accelerated iterative x-ray CT image reconstruction using floating point graphics hardware. 1984. Evaluation of the ordered subset convex algorithm for cone-beam CT. 50(4):613– 623. Kole and F. 34(10):3732–3738. A. MRF Modeling in Computer Vision. November 1993. Stierstorfer. 1992. H¨rer. L. Link¨ping. and J.Bibliography 163 [52] J. 1951. R. Utah.de/phantoms. IEEE Transactions on Image Processing. An Iteration Formula for Fredholm Integral Equations of the First Kind. American Journal of Mathematics. Li. S. In 8th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. 1998. Kole and F. Lewitt. US Patent 5. [54] H. 51(4):875–889. 2006. Lauritsch and H. Fessler. Resolution and noise trade-off analysis for volumetric CT. 4(10):1430–1438. Medical Physics. Springer Verlag. 37(3):705–716. EM reconstruction algorithms for emission and transmission tomography. Salt Lake City. [62] B. 1995. Journal of Computer Assisted Tomography. As available electronically in May 2009 at http://www. Beekman. Nutating slice CT image reconstruction apparatus and method. M. Head phantom. and C. [59] G. 2005. Globally convergent algorithms for maximum a posteriori transmission tomography. Kunze. o o [65] M. [60] R. Physics in Medicine and Biology. Avinash. and J. C. Lange and R. C. Larson. 2006. Danielsson. Magnusson. Physics in Medicine and Biology. . 1995.134. Link¨ping University.802. Department of Electrical Engineering. Sunneg˚ ardh. PhD thesis 320. and W.

Mortelmans. Lindau. Anti-aliased three-dimensional conebeam reconstruction of low-contrast objects with algebraic methods. [74] NVIDIA Corporation. Mueller and R. 2007. Discrete Radon transform in a continuous space. Journal of the Optical Society of America A. M. J. J. IEEE Transactions on Medical Imaging. 43(4):729–737. O’Sullivan. R. Nocedal and S. Honda. Springer. M. [78] T.1. [73] J. version 2. [71] J. NVIDIA CUDATM programming guide. [79] A. P. Nassi. version 1. 28(4):3641–3647. Peters. In 9th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. and J. Ohyama.nvidia. and L. 1999. Tsujiuchi. Suetens.164 Bibliography [66] B. second edition. Mueller. Defrise. Polacin. IEEE Transactions on Medical Imaging. Stuttgart. Medoff. Germany. 4(4):200– 207. S. Wheller. J. 18(6):519–537. IEEE Transactions on ASSP.and re-projection in computed tomography. R. Physics in Medicine and Biology. Natterer. [75] NVIDIA Corporation. IEEE Transactions on Nuclear Science. Yagel. The mathematics of computerized tomography. Numerical optimization. Wheller. [69] K. 1985. 1981. Dupont. [77] J. Rapid 3-D cone-beam reconstruction with the simultaneous algebraic reconstruction technique (SART) using 2-D texture mapping hardware. [68] K. Fast implementations of algebraic methods for three-dimensional reconstruction from cone-beam data. and J. Mueller. B. Brody. As available electronically in May 2009 at http://www. 1983. and A. Yagel. S. Inoue. Some windows with very good sidelobe behavior. R. As available electronically in May 2009 at http://www. 1986. Wiley and B. P. J.com. Wright. [72] A. Yagel.1. De Man. W. Algorithms for fast back. Iterative convolution backprojection algorithms for image reconstruction from limited data. Iterative reconstruction for helical CT: a simulation study.G. and T. NVIDIA CUDATM programming guide. . 1998.com. 1981. Ohki. [76] N. Springer Series in Operations Research. [70] F. 1999. Nuttall. 73(11):1493–1500. A fast sinc function gridding algorithm for fourier inversion in computer tomography. [67] K. Journal of the Optical Society of America. 19(12):1227– 1237. IEEE Transactions on Medical Imaging. 1999. Nuyts. P. 18(6):538–548. IEEE Transactions on Medical Imaging. Teubner. M. 1987.nvidia. 29(1):84–91. Macovski. D. Design consideration on image reconstruction system for highend CT-scanner. 4:318–324. 2000.

H. Th. editors. 21(1):228–236. N. M H Bourguignon. Rodrigues. B. A.-M. Fessler. M. volume 97. Grass. Presented at the First Workshop on General Purpose Processing on Graphics Processing Units. and J P Kernevez. Hsieh and M. Proksa. J¨hne. M. Radon. 1982. Kohler. S. 68(9):2236–2240.M. Numerische Isotropieoptimierung o a von FIR-Filtern mittels Quergl¨ttung.A. Baghsorkhi. Keck. [91] H. K¨rkel. Braunschweig. In E. F. Medical Physics. M. IEEE Transactions on Medical Imaging. W. Quadratic Regularization Design for Iterative Reconstruction in 3D multi-slice Axial CT. volume 6510. Proceedings of the National Academy of Sciences of the United States of America. Mustererkennung. Kohler. Stone. editor. volume 5. Radon. T. In Eric C. Proksa. 1(2):113–122. a [82] G. 2004. M. Implementation of the FDK Algorithm for Cone-Beam CT on the Cell Broadband Engine Architecture. Scherl. edia tors. Vienna. 2000. The frequency split method for helical cone-beam reconstruction. Wahl. Kowarschik. Birkh¨user Verlag. and J. The n-PImethod for helical cone-beam CT. [90] L. and W. pages 4464–4466. 1971. Paulus and F. Shechter. Scharr. 1917. [83] C Riddell. Physics in Medicine and Biology. and B. [87] H. Reconstructing interior head tissue from x-ray transmissions. IEEE Nuclear Science Symposium Conference Record. volume 6. and R. S. Logan. Frey. Lakshminarayanan. Ueng. B Bendriem. In J. IEEE Transactions on Nuclear Science. 2007. Berichte S¨chsiche Akademie der Wisa a a senschaften. 2006. Eckert. Reprinted in J. ¨ [81] J. Shepp and B. [84] S. . In IEEE Nuclear Science Symposium Conference Record. Flynn. [88] G. S. Proksa. Timmer. Kowarschik. Shi and J. Proceedings of SPIE. Scherl. 1974. Fast GPU-Based CT Reconstruction using the Common Unified Device Architecture (CUDA). Ryoo. Program Optimization Study on a 128-Core GPU.-Z. Uber die Bestimmung von Funktionen duch ihre Integralwerte l¨ngs gewisser Mannigf¨ltigkeiten. A. 1995. 69:262–267. [89] L. Altman. pages 367–374. Maximum likelihood reconstruction for emission tomography. R. September 1997. Hofmann. C. October 2007. V. and J. Hornegger. Springer. 31(8):2230– 2236. 19(9):848–863. Three-dimensional reconstruction from radiographs and electron micrographs: application of convolutions instead of Fourier transforms. The approximate inverse and conjugate gradient: non-symmetrical algorithms for fast attenuation correction in SPECT. Gesammelte Abhandlungen. Hornegger. Vardi. S. IEEE Transactions on Medical Imaging. [86] H. and J. Shepp and Y.Bibliography 165 [80] R. pages 2834–2836. Hwu. Koerner. [85] H. 1987. J. 40(2):269. Ramachandran and A. A. 2007.

H. I. L. 6(3):228–238. M. Stierstorfer. 49(11):2209–2218. o [103] J. M. Schaller. pages 1918–1927.-E. Kazama. University of Michigan. 1985. [93] E. Miller. J. L. [99] S. Master’s thesis.uni-erlangen. volume 5032 III. Link¨pings Universitet. 2008. Germany. [98] S. Thomas Jr. Noise and edge artifacts in maximum-likelihood reconstructions for emission tomography. June 2008.-E.. Sunneg˚ ardh. Iterative improvement of non-exact reconstruction in cone-beam CT. Sunneg˚ ardh and P. Mori. A new anti-aliased projection operator for iterative CT reconstruction. pages 5426–5433. Salt Lake City. Sunneg˚ ardh. Knaup. 12(2):252–255. [96] D. Thorax phantom. Flohr. [97] K. September 2004. Physics in Medicine and Biology. 2007. February 2007. [100] K. 1985. S.The International Society for Optical Engineering. Weighted FBP–a simple approximate 3D FBP algorithm for multislice spiral CT with good dose usage for arbitrary pitch. I. Windmill artifact in multi-slice helical CT. Sidky and X. A. Taguchi. [101] J. Miller. The use of sieves to stabilize images produced with the EM algorithm for emission tomography. total-variation minimization. Ann Arbour. Snyder. Srivastava. 32(5):3864–3872. 2004. Medical Physics. 2008. In 9th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. Licentiate Thesis.de/phantoms. Boese. I. L. Link¨pings Universitet. IEEE Transactions on Nuclear Science. Sunneg˚ ardh and P. 2003. Combining analytical and iterative reconstruction in helical cone-beam CT. G. Siddon. Chiang. IEEE Transactions on Medical Imaging. and D. Politte. Phd thesis. Lewis J. B.166 Bibliography [92] R. Kachelriess. Danielsson. [95] D. [94] M. Silver. Hyperfast general-purpose cone-beam spiral backprojection with voxel-specific weighting. and I. Steckmann. In IEEE Nuclear Science Symposium Conference Record. Danielsson. Image reconstruction in circular cone-beam computed tomography by constrained. M.imp. In 8th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. Utah. In Proceedings of SPIE . K. Physics in Medicine and Biology. Accelerated statistical image reconstruction algorithms and simplified cost functions for x-ray computed tomography. USA. and T. Rauscher. 2005. Snyder and M. . Lindau. As available electronically in May 2009 at http://www. Bruder. [104] J. 53(17):4777–4807. Hein. Iterative enhancement of non-exact reconstruction in cone beam CT. o [102] J. Y. and M. Pan. Sourbelle. Fast calculation of the exact radiological path for a threedimensional CT array. 1987.

Bouman. Turbell. pages 571–574. 2007. [116] F. Sweden. 35(9):4173–4185. Lindau. 1993. A.-S. volume 1. 5(1):89–94. . Bristol. W. From the watching of shadows : the origins of radiological tomography. 6(4):313–319.Bibliography 167 [105] J.-E. 52(6):R1–R13. Approximate and exact cone-beam reconstruction with standard and non-standard spiral scanning. 20(6):1675– 1684. Sunneg˚ ardh and P.775. 2008. Department of Electrical Engineering. Mueller. Danielsson. Grasruck.112. 1987. and H.-E. 2007. Sunneg˚ ardh and P. PhD thesis 672. Ultra-fast 3D filtered backprojection on commodity graphics hardware. Physics in Medicine and Biology. February 2001. Sauer. Chesler. August 2000. Ye.-L. Veklerov and J. IEEE Transactions on Medical Imaging. 1981. K. Correia.D. Liow. Iterative algebraic reconstruction algorithms for emission computed tomography: A unified framework and its application to positron emission tomography. Attenuation correction in gamma emission computed tomography. [115] S. Medical Physics.-B. Danielsson. Regularized iterative weighted filtered backprojection for helical cone-beam CT. Sunneg˚ ardh and M. and J. Xu and K. Link¨ping University. pages 5090–5095. [110] H. o Link¨ping.390. and S. and J. Hilger. Medical Physics. [114] G. Medical Physics. Yu. Germany. [109] J. Webb. C. Nonlinear regularization of iterative weighted filtered backprojection for helical cone-beam CT. In 9th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. 2004. [112] E. Cone-Beam reconstruction using filtered backprojection. A three-dimensional statistical approach to improved image quality for multislice helical CT. Three-dimensional computerized tomography scanning method and system for large objects with smaller area detectors. A. Llacer. C. In IEEE International Symposium on Biomedical Imaging: Nano to Macro. Y. Xu. Walters. Stopping rule for the MLE algorithm based on statistical hypothesis testing. E. J. [117] X. Regularized iterative weighted filtered backprojection for helical cone-beam CT. K. Thibault. In IEEE Nuclear Science Symposium and Medical Imaging Conference Record. 34(11):4526–4544. 1990. February 1995. D. 3D image reconstruction for helical partial cone beam scanners using wedge beam transform. 2007. [108] K.A. o [111] H. Hsieh. [113] T. [107] J. US Patent 6. Wang.104. [106] J. C. Journal of Computer Assisted Tomography. Simon. Tam. 2008. Strother. Tuy. US Patent 5.

Firmin. [123] K. Z. T. Nielsen. 19(5):548–555. D. [120] A. Hein. 34(2):585– 598. USA. An error-reduction-based algorithm for cone-beam computed tomography. Nielsen. and R. [119] A. and M. Chen. In IEEE Nuclear Science Symposium Conference Record. K¨hler. I. Efficient projection and o backprojection scheme for spherically symmetric basis functions in divergent beam geometry. Partial volume and aliasing artefacts in helical cone-beam CT. M. L. and R. 33(12):4653–4663. A. [121] W. [129] Y Zou. Ziegler. Proksa. Grass. [127] A. Yang and D. Proksa. and R. 2004. pages 2293–2295. Gullberg. 2000. and S.168 Bibliography [118] G. [122] G. Exact image reconstruction on PI-lines from minimum data in helical cone-beam CT. A. Silver. Zbijewski. 2006. N. 2005. 2004.L. Medical Physics. IEEE Engineering in Medicine and Biology Magazine : the Quarterly Magazine of the Engineering in Medicine & Biology Society. Katsevich. and G. [126] A. T. pages 2264–2267. 2006. IEEE Transactions on Medical Imaging. Utrecht. T. 2006. Juni 2006. Utrecht University. T. Zamyatin. Y. Z. Up-sampling with Shift Method for Windmill Correction. Helical CT reconstruction with large cone angle. Ziegler. and S. Nielsen. Zhang. A.T. T. Proksa. 2007. T. [128] Y. Model-based image reconstruction in X-ray Computed Tomography. Iterative reconstruction of a region of interest for transmission tomography. [125] A. Nakanishi. Wang. Ziegler. Medical Physics. 49(6):941– 959. Phd thesis. . Zamyatin. 2008. Medical Physics. 2004. 19(1):120–125. E. 31(12):3206– 3212. volume 4. D. Physics in Medicine and Biology. Nakanishi. 49(11):2365–2375. In 8th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. The birth of the first CT scanner. Ziegler. Noise and resolution in images recono structed with FBP and OSC algorithms for CT. Sidky. Pan. Unmatched projector/backprojector pairs in an iterative reconstruction algorithm. [124] A. Utah. Salt Lake City. Zeng. volume 4. Pan. Physics in Medicine and Biology. K¨hler. Zeng and G. A. Zou and X. 35(4):1317–1327. M. 2000. Medical Physics. In IEEE Nuclear Science Symposium Conference Record. K¨hler. Silver. and X. Iterative cone-beam CT o image reconstruction with spherically symmetric basis functions.

Sign up to vote on this title
UsefulNot useful