This action might not be possible to undo. Are you sure you want to continue?
Link¨ oping Studies in Science and Technology
Dissertation No. 1264
Iterative Filtered Backprojection
Methods for Helical ConeBeam CT
Johan Sunneg˚ardh
Department of Electrical Engineering
Link¨ opings universitet, SE581 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
SE581 83 Link¨ oping, Sweden
Copyright c ( 2009 Johan Sunneg˚ardh
Sunneg˚ardh, Johan
Iterative Filtered Backprojection Methods for Helical ConeBeam CT
ISBN 9789173935869
ISSN 03457524
Typeset with L
A
T
E
X
Printed in Sweden by LiUtryck, Link¨ oping, 2009
iii
Abstract
Stateoftheart reconstruction algorithms for medical helical conebeam Com
puted Tomography (CT) are of type nonexact Filtered Backprojection (FBP).
They are attractive because of their simplicity and low computational cost, but
they produce suboptimal 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 nonexact 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 nonexactness 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 (WLSIFBP). 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 edgepreserving nonlinear reg
ularization. In this case, resolution becomes contrast dependent, a fact that can
be utilized for improving high contrast resolution without degrading the signalto
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 nonlinear regularization on the other
hand, the degraded signaltonoise 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 ickeexakta. 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 ickeexakta metoden WFBP [Stierstorfer et al., Phys. Med.
Biol. 49, 22092218, 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 OSIWFBP. 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 WLSIFBP, som
med en gradientbaserad metod minimerar en kvadratisk m˚alfunktion. Den nya
metoden undertrycker artefakter ungef¨ ar lika eﬀektivt som RIWFBPmetoden,
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 ickelinj¨ ar regularisering unders¨ okts, vilket betyder att de rekon
stuerade bilderna i varje iteration ﬁltreras med ett ickelinj¨ 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/brusf¨ 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 WLSIFBP 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 ickelinj¨ 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, PerErik 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
proofreading, 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 proofreading, 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 proofreading 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
proofreading 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 Twodimensional 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 (OSIWFBP) 71
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2 The Ordered Subsets (OS) technique . . . . . . . . . . . . . . . . . 71
6.3 OSIWFBP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 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 (WLSIFBP) . . . . . . . . . . . . . 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 nondestructive 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 Xray images acquired in multiple di
rections to calculate crosssectional images of the Xray 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 Xray 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 zaxis, 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 zaxis. Its height h when projected onto the center of
rotation, or equivalently, the coneangle 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
Xray 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 conebeam 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 nonexact. 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 nonexact 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, nonexact methods are computationally less de
manding, simpler to implement, and oﬀer a better dose utilization than exact
methods.
One stateoftheart analytical nonexact 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).
conebeam 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. Windmill artifacts are caused by aliasing in
the zdirection of the detector, and look like alternating dark and bright streaks ex
tending radially from the upper and lower ends of the spheres. The coneartifacts
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 nonexact
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 nonexact reconstruction in helical conebeam 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 wellknown 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 (OSIWFBP). Ordered Subsets (OS)
is a wellknown 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 OSIWFBP 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 (WLSIFBP) method. Evaluation by visual inspec
tion, noise measurements, and root mean squared error measurements show that
WLSIFBP produces images very similar to those of RIWFBP, but with a slightly
slower rate of convergence than for RIWFBP. In return, the WLSIFBP 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., nonquadratic potential functions, can be used with the RIWFBP
and WLSIFBP. This is followed by experiments with WLSIFBP and the q
Generalized Gaussian Markov Random Field (qGGMRF) potential function by
Thibault et al. [109]. Noise and spatial resolution measurements at diﬀerent con
trast levels show that nonlinearly regularized WLSIFBP can be used to improve
high contrast spatial resolution and reduce artifacts while increasing the signal to
noise ratio for lowcontrast 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 WLSIFBP 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 scalarvalued 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 Twodimensional image reconstruction
The mathematical problem of twodimensional 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 xray 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 twodimensional 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 twodimensional function:
10 Analytical reconstruction methods
¦0 )=
1
I
0
¦ v)al
v
I
0
v
Figure 2.1: Photon intensities of attenuated and nonattenuated xray 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 twodimensional Fourier trans
form Tf of a function f(x, y) along a radial line equals the onedimensional
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
twodimensional Fourier transform of f(x, y) through a series of onedimensional
Fourier transforms T
t
({f). Performing a ﬁnal inverse twodimensional 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 Twodimensional 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 nonzero 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 onedimensional 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 lowpass 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
2Δ
2
t
sinc(t/Δ
t
) −
1
4Δ
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
4Δ
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 nonzero.
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
]Λ
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 conebeam 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 conebeam geometry. Right: the semiparallel
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 conebeam data to semiparallel data (Fig. 2.5),
2) Rampﬁltering 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) =
Pα
2π
,
2.2 Weighted ﬁltered backprojection 15
Xray source
Figure 2.6: Scanning geometry in the xyplane.
P
Xray source ( ÷ 0)
Detector
max
1
0
1
D
Figure 2.7: Scanning geometry in the yzplane.
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 xyplane,
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 xyplane, the rebinned
projections look perfectly parallel. However, the rays diverge in the zdirection
and the rebinned projections are therefore called semiparallel 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 lowpass 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 twodimensional 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 nonzero 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 nonlocal 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 xray 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 conebeam 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 SignaltoNoise Ratio (SNR), a fact which has forced the CT
industry to rely on nonexact 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 conebeam
CT [73, 16, 121, 127, 109, 98].
• Incomplete data. One example of how to suppress artifacts due to missing
data in circular conebeam 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 xrays 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 nonexactness
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 twodimensional array of image coeﬃcients are
mapped onto a continuous function by interpolation with a bilinear basis function.
The sampling distance in the ydirection is assumed to be equal to the sampling
distance in the xdirection.
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 threedimensional 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 threedimensional case, if nonlinear eﬀects are ne
glected, the contribution from a continuous threedimensional 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 threedimensional 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 costeﬀective 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 DCcomponent will appear in the result, but also
false high frequency components. This aliasing of neighboring DCcomponents
was named “DCaliasing” 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
(Δ
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 nonzero 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 wellknown example of rotationally symmetric basis functions satisfying (i)
and (ii) above is the class of generalized KaiserBessel 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
(Δ
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 xray 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 zdirection after
semiparallel 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
0º
45º
2
2
RemagDCcompoet
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
Nondivergent 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
antialiasing 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 DCaliasing. 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
2Δ
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 noncentral repetitions of the DCcomponent 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 wellbehaved
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) illposedness of the tomography problem
2
and (ii) insuﬃcient restrictions on the space of possible solutions.
In the context of tomographic reconstruction, illposedness 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 illposedness 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 wellbehaved 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 KaiserBessel
basis functions (also known as “blobs”) for iterative reconstruction. The amplitude
2
A problem is said to be illposed 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 suboptimal 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 ) = βf 
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 zdirections
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
Noisefree 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
submeasurements. 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 nonattenuated photons traveling along the ray i from (4.1).
Then, the number of attenuated photons I is approximately Poisson distributed
36 Simulation and evaluation
Xray 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 subcontributions.
Table 4.1: Scanning parameters for experiments in following chapters.
SourceIsocenter distance R
S
595mm
IsocenterDetector 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 xyplane N
SW
3
Source subsampling in zdir N
SL
3
Angle subsampling N
GR
3
Active detector area (xyplane) d
β
0.8
Active detector area (zdir) 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
Tablefeed 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 TamDanielsson
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 semiparallel geometry. In the rebinning process, new parameters de
scribing sampling in the new semiparallel 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 zdirection. 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 noisefree 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 inplane (xyplane) spatial resolution, repre
sented by Modulation Transfer Functions (MTFs), and crossplane (zdirection)
spatial resolution, represented by Slice Sensitivity Proﬁles (SSPs).
The most direct and common way of measuring the inplane MTF is to use a
thin wire test object directed orthogonal to the xyplane [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 zaxis is often used to measure the
crossplane 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 edgebased 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 stepbystep 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 inplane
42 Simulation and evaluation
v
s
0
s
1
s
N
r
s
0
s
N
Edge response
Binning
r
t
1
t
N1
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 crossplane 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 abovementioned 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 nonlinear 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 nonexact 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 nonexact 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 ﬁnitedimensional 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 xray 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 semiparallel 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 preprocessed 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
ﬁ
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 lowpass 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
+β
z
R
z
)/C
N
, (5.6)
where β
xy
and β
z
are parameters used for tuning the inplane and crossplane
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 lefthand side matrices corresponds
to the convolution operations on the righthand 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 nonexact, nonlocal 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 noniterative and iterative results. Besides aliasing
artifacts known from twodimensional 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 highfrequent 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 lowfrequency 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 closeups 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 zdirection. 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 preﬁltered
f
5
image from the overshoots in Fig. 5.7 b). Preﬁltering 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 zaxis. 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 isocenter and 140mm away from the
isocenter. Regularization parameters: β = 0.52, β
xy
= 1.0, β
z
= 1.5. The WFBP
and RIWFBP curves overlap both in the isocenter and oﬀcenter 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 dashdotted 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 nonregularized 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 PImethod [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 PImethod, incorporates data
outside the TamDanielsson 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 nonexact 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 preprocessing suppresses artifacts nonlocal
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 nonstatistical 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 nPI 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 windmill artifacts in helical
conebeam CT. Software approaches include interpolation between conjugate rays
during backprojection [38], nonlinear 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 3PI detector for a pitch factor of 0.75. White: 3PI detector.
Light gray: clinical CT detector. Dark gray: PIdetector.
alleviating windmill 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 windmill artifact reduction, there also exists a hardware technique already
employed in clinical CT systems [27]. By rapidly moving the focal spot between
two diﬀerent zpositions, 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 windmill artifacts. Unfortunately, the moving spot mechanism comes with
increased complexity and costs in many parts of the acquisition system. Therefore,
software solutions for alleviating windmill 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 highpass 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 nonlinear 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, nonlinear regularization has shown to be useful for preserving highcontrast
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 columnaction 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 nonexistent.
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
(OSIWFBP)
6.1 Introduction
This chapter is an adapted and reduced version of [102, Chapter 6], presenting
the OSIWFBP method, which is a nonregularized and accelerated version of the
RIWFBP method. Since the original publication, our investigations have focused
on other topics, mainly because the OSIWFBP method seemed to be useful only
for extremely high cone angles (κ
max
> 4.8
◦
), which are unlikely to be used for
medical helical conebeam 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 OSIWFBP method. This is followed by an experiment
showing how OSIWFBP 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 (OSIWFBP)
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 OSIFBP 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
+α
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 θoﬀset (¦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 OSIWFBP 73
6.3 OSIWFBP
Similarly to the RIWFBP, the ﬁrst step in the OSIWFBP method is rebinning of
the input data p
in
to semiparallel 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 OSIWFBP is given by
f
k+1
= f
k
+α
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 subiterations.
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, rowwise lowpass ﬁltering
of projection data diﬀerences (p
reb
−P
i
k
f
k
) is incorporated in the reconstructions
Q
l
. The kernel used for lowpass 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 DCcomponent 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 twodimensional 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 threedimensional 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 noiselike artifacts appear in the OS
IWFBP results (see Fig. 6.6 c)). As a numerical measure of these artifacts, the
74 Ordered Subsets IWFBP (OSIWFBP)
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 OSIWFBP.
SourceIsocenter 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
Tablefeed 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 σvalues determining the amount of
lowpass 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 OSIWFBP 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 64row
system with a maximum cone angle κ
max
= 4.8
◦
and one 128row system with
76 Ordered Subsets IWFBP (OSIWFBP)
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 OSIWFBP 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 lowpass ﬁltering.
78 Ordered Subsets IWFBP (OSIWFBP)
a) RIWFBP, α = 1.0, β = 0 b) OSIWFBP, α = 0.15, No LP
c) OSIWFBP, α = 0.25, No LP d) OSIWFBP, α = 0.25, σ = 1
Figure 6.6: Blown up twodimensional reconstructions of the Forbild Thorax
phantom [97], showing highfrequency “ringing” artifacts, and how they can be
suppressed by the lowpass 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 OSIWFBP 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 OSIWFBP
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 OSIWFBP, 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
+α
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 OSIWFBP eﬃciently improves the recon
struction results during the ﬁrst iterations. However, since there is no mechanism
to control the step lengths of the method, OSIWFBP is not convergent. Instead,
80 Ordered Subsets IWFBP (OSIWFBP)
(a) RIWFBP, β = 0, f
0
(b) OSIWFBP, f
0
(c) RIWFBP, β = 0, f
1
(d) OSIWFBP, f
10
(e) RIWFBP, β = 0, f
2
(f) OSIWFBP, 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 OSIWFBP 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) OSIWFBP, f
0
(c) RIWFBP, β = 0, f
1
(d) OSIWFBP, f
10
(e) RIWFBP, β = 0, f
2
(f) OSIWFBP, 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 (OSIWFBP)
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 OSIWFBP
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 WLSIFBP, 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 matrixvector 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 nonnegative. There
fore, (7.6) is satisﬁed if and only if the matrix SQP can be written as VBP, where
BP is symmetric and nonnegative. In such case, a factorization BP =
ˆ
P
T
W
ˆ
P
exist, where W is a symmetric nonnegative 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 (WLSIFBP) method, which is similar to the RIWFBP method but has an
update step that satisﬁes (7.5).
7.2 Weighted least squares IFBP (WLSIFBP)
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 (WLSIFBP) 85
we consider the problem of ﬁnding a vector/reconstructed image f minimizing the
function
z(f ) =
1
2
Pf −p
reb

2
W
+β
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 WLSIFBP 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 lowfrequency artifacts when applied as a di
rect reconstruction operator. Therefore, in order to reduce the number of required
iterations, the WLSIFBP 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 nonnegative, 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) WLSIFBP, f
5
Backprojector: P
T
b) WLSIFBP, f
5
Backprojector: B
S
Figure 7.1: WLSIFBP 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
WLSIFBP using noisefree data. In any forward projection or backprojection
operation, the interpolation function must be adapted to the input sampling grid
in order to avoid DCaliasing [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 noniterative 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 nonexactness of the reconstruction method,
and (ii) windmill artifacts showing up as alternating bright and dark regions near
sharp edges orthogonal to the zdirection of the reconstruction volume.
Fig. 7.2 c)f) show results after one and two iterations with the RIWFBP and
WLSIFBP 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 WLSIFBP 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 WLSIFBP 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 noniterative WFBP results in strong nonlocal
artifacts. Again, the RIWFBP and WLSIFBP 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 WLSIFBP 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 WLSIFBP 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
WLSIFBP are nearly identical irrespective of distance to isocenter.
As shown in Fig. 7.7, also the SSPs for RIWFBP and WLSIFBP are very
similar. In contrast to the xyplane 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 WLSIFBP at approximately equal spatial resolution. Again, there
are only small diﬀerences between the RIWFBP and WLSIFBP 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) WLSIFBP, f
1
e) RIWFBP, f
2
f) WLSIFBP, f
2
Figure 7.2: RIWFBP and WLSIFBP 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) WLSIFBP, f
5
c) RIWFBP, f
5
d) WLSIFBP, f
5
Figure 7.3: RIWFBP and WLSIFBP 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 WLSIFBP 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 WLSIFBP 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 WLSIFBP 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) WLSIFBP, f
5
, Q = 0.7 f) WLSIFBP, f
5
, Q = 1.0
Figure 7.4: Comparison between WFBP, RIWFBP, and WLSIFBP 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
)
WLSIFBP,Q=0.3 WLSIFBP,Q=0.7 WLSIFBP,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
WLSIFBP, f
5
, r = 0mm
RIWFBP, f
5
, r = 150mm
WLSIFBP, f
5
, r = 150mm
Figure 7.6: MTF measurements of the RIWFBP and the WLSIFBP method in
the isocenter and 150mm oﬀcenter. The RIWFBP and WLSIFBP curves overlap
both in the isocenter and oﬀcenter 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
WLSIFBP, f
5
, r = 0mm
RIWFBP, f
5
, r = 150mm
WLSIFBP, f
5
, r = 150mm
Figure 7.7: Normalized SSP measurements of the RIWFBP and the WLSIFBP
method in the isocenter and 150mm oﬀcenter.
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
WLSIFBP, Q = 0.3
RIWFBP, Q = 0.7
WLSIFBP, Q = 0.7
RIWFBP, Q = 1.0
WLSIFBP, 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 nonlinearities 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 WLSIFBP. The next section presents nonlinearly regularized ver
sions of the RIWFBP and WLSIFBP methods, and so called potential functions,
employed to deﬁne the nonlinearity 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 WLSIFBP methods,
we consider the following cost function
z
V
(f ) =
1
2
Pf −p
reb

2
W
+β
¸
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 nonconvex and convex potential functions have been suggested
for image reconstruction and restoration. Examples of nonconvex 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 qgeneralized Gaussian Markov Random ﬁeld (qGGMRF) 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 qGGMRF 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 qGGMRF
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 vshaped, 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 qGGMRF
and the quadratic function should occur.
The qGGMRF 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: qGGMRF 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
ﬂ
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: qGGMRF 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 WLSIFBP and
RIWFBP were found to be very similar. Assuming this similarity to be valid
also for nonlinear regularization we limit the experiments in this chapter to the
WLSIFBP 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 preﬁltered 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 preprocessed 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 crossplane 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 inplane 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 inplane 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 nonregularized case. As the contrast level increases to 500HU, the relative
amount of low contrast smoothing becomes smaller, and the SSP approaches that
of the nonregularized method.
As shown already in Fig. 8.3, the number of iterations needed for convergence
of crossplane 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 WLSIFBP(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 nonnormalized 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
noisefree data. The image is corrupted by streak artifacts plus a combination of
coneartifacts, windmill 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 WLSIFBP 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 WLSIFBP, we have studied head
phantom reconstructions for diﬀerent numbers of nonattenuated 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
/σ
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 WLSIFBP method, for which the above presented proportionality holds.
In relation to the linearly regularized WLSIFBP method, the nonlinearly regu
larized WLSIFBP 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 WLSIFBP, and nonlinear WLSIFBP. The WFBP recon
struction is corrupted by windmill artifacts. As shown in Fig. 8.13 b), the linear
WLSIFBP method gives a better result with less pronounced windmill artifacts.
The nonlinear WLSIFBP in Fig. 8.13 c) further reduces the amplitude of these
artifacts, as well as improves the crossplane resolution as can be readily seen in
the encircled areas.
8.3 Experiments 105
a) Sampled phantom b) WFBP reconstruction
c) WLSIFBP, f
10
, p = q = 2 d) WLSIFBP, f
10
,
p = 2.0, q = 1.3, c = 20
Figure 8.10: Axial images of noisefree 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: WLSIFBP 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
/σ
nq
against the number of unattenuated photons. σ
nq
are noise
measurements for WLSIFBP 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) WLSIFBP, f
10
, p = q = 2.0
c) WLSIFBP, 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 windmill artifacts in the WFBP reconstruc
tion, and how the amplitude of these to a certain extent can be reduced with the
linear WLSIFBP method. For the nonlinear WLSIFBP method, two values of β
were used. As shown in Fig. 8.14 c), the nonlinear WLSIFBP with β = 0.6 does
not improve the result much over the linear WLSIFBP. 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 WLSIFBP reconstruction with respect to windmill 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 lowpass ﬁltering S introduced in Section 5.1. As
illustrated by the simple onedimensional 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 suboptimums 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) WLSIFBP, f
10
, p = q = 2.0
c) WLSIFBP, f
10
, q = 1.3, c = 50, β = 0.6
d) WLSIFBP, 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: Onedimensional 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 qGGMRF 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 WLSIFBP 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 viewdependent 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 viewindependent trilinear 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 conebeam data domain, it could be possible to avoid rebinning entirely, but
this has not been investigated so far. Instead, we accept the lowpass ﬁltering
caused by rebinning, and focus on improving the reconstruction by modeling the
lowpass ﬁltering in the acquisition process.
We start with a description of the improved conebeam 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
2π
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 zaxis. 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 xyplane, the detector elements face the source as shown in Fig. 9.1.
Furthermore, the detector elements are aligned to the zaxis 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
Xray source Detector element
R
SD
R
S
l
S
]cos¦nn)]
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 xyplane. Note that the sizes of the source and detectors are
exaggerated.
SC
DC
z
x
SL
DH
Xray 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 xzplane. 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
(α
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
(α
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 subangles 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 isocenter and
one located 150mm oﬀcenter for measuring oﬀcenter 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 Nvalue (e.g. N
SL
) should be set to a value
larger than one. The higher the Nvalue 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 Nvalues quickly
leads to prohibitively high computational costs. In the experiments, the Nvalues
have been empirically selected such that any further increase in Nvalue 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, DCaliasing artifacts may occur when using P
T
corresponding to the
Joseph method with only one ray per measurement.
Interestingly, DCaliasing 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 DCaliasing 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 WLSIFBP 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 DCaliasing artifacts occur mainly because the in
terpolation function in the zdirection 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 DCaliasing 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 DCaliasing
artifacts.
We begin with studying crossplane resolution in the isocenter, 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 crossplane resolution.
9.2 Multiple ray forward projection 119
a) WFBP reconstruction b) WLSIFBP, f
10
, B
S
N
SL
= 1, N
SW
= 1
N
DH
= 1, N
DW
= 1
c) WLSIFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 1, N
DW
= 1
d) WLSIFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 3, N
DW
= 1
e) WLSIFBP, f
10
, P
T
N
SL
= 1, N
SW
= 1
N
DH
= 3, N
DW
= 2
f) WLSIFBP, 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 WLSIFBP.
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.
Inplane resolution can be divided into radial resolution in the isocenter, and
radial and azimuthal (angular) oﬀcenter 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 crossplane resolution case, the
resolution does not continue to improve after the ten ﬁrst iterations. Interestingly,
in contrast to the crossplane 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 oﬀcenter 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 WLSIFBP after ten iterations. Left: complete SSP. Right:
Blowup 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: isocenter MTF ρ
50%
values for WLSIFBP. 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: oﬀcenter radial MTF ρ
50%
values for WLSIFBP. 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 oﬀcenter radial resolution, and
has not been modeled here, i.e. N
GR
= 1.
an object located 150mm oﬀcenter. This object is irradiated at a maximum fan
angle of β = sin
−1
(150mm/R
S
). Ignoring the width of the source, at this fanangle
the object “sees” the source as if it is l
S
150mm/R
S
≈ 3mm wide. Clearly, the
width of the resulting xray beam will cause blurring in the radial direction.
The left of Fig. 9.9 shows how oﬀcenter 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 oﬀcenter 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: oﬀcenter azimuthal MTF ρ
50%
values for WLSIFBP. 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 crossplane 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 crossplane resolution (indicated
with solid circles). Unfortunately, another obvious and immediate consequence of
the detector element modeling is an increased amount of windmill artifacts.
Fig. 9.12 shows how nonlinear regularization can partially be used to control
the increased amount of windmill artifacts. The upper row, i.e. Fig. 9.12 a)
and b), shows axial images obtained with linear and nonlinear 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 crossplane resolution is further improved by the detector
element modeling. The P
T
result has a slightly higher crossplane resolution,
but as indicated by the dashed circles, this image also contains more windmill
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: WLSIFBP 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
inplane 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 crossplane resolution than B
S
, and B
S
gives higher inplane
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 WLSIFBP 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, inplane
−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, inplane
−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, crossplane
−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, crossplane
Figure 9.13: Interpolation functions and their Fourier transforms for B
S
and
P
T
in the tdirection and qdirection 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 lowpass 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
oﬀcenter 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 inplane resolution is limited by initial rebinning
of input data. One possibility to further improve inplane resolution could be
to entirely avoid rebinning by performing forward projection and backprojection
directly in the conebeam geometry, and use a rampﬁlter derived for fanbeam
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 inplane 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 DCaliasing
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 2D parallel beam
CT. Except for some small details, generalization to 3D conebeam 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 Pmatrix 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 Pmatrix 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
(θ
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
(Δ
x
ζ(θ
i
)ρ) sinc(Δ
t
ρ). (9.12)
In order to avoid DCaliasing during backprojection, H
i
(ρ) must be zero for ρ =
kΔ
−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, 2D reconstructions of the Shepp
Logan phantom [89] have been studied. Noisefree input data were generated from
a SheppLogan phantom with a length of 368mm. Each detector contribution was
calculated as a mean value of seven parallel subrays. 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
+β
N
¸
i=1
N
¸
j=1
d
ij
(f
i
−f
j
)
2
2
. (9.13)
This is equivalent to the WLSIFBP 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 SheppLogan 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
/Δ
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
/Δ
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 DCaliasing 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 parallelbeam
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 semiparallel 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 SheppLogan 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 SheppLogan 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 conebeam 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 conebeam geometry where the adaptive antialiasing
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
+Δ
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 precalculated
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 nonzero 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 zdependent angle φ = φ
0
+ 2πz/P. These techniques are very
eﬃcient. Furthermore, since all coeﬃcients are precalculated, 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 lookup tables, the backprojected images must be
136 Implementation on Graphics Processing Units (GPU)
resampled to a Cartesian grid, and lookup 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 FieldProgrammable
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 wellknown 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 400600
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 readonly
region of the global GPU memory. The texture cache improves memory access
from 1D, 2D, and 3D 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 1D, 2D, or 3D 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 forloop.
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  ifstatement 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
ifstatement 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 1D, 2D, and 3D 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 2D
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 3Dtexture to it.
2 Conﬁgure grid and thread blocks.
3 Calculate constant memory lookup tables for simpliﬁed computation
4 of the xray 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 threedimensional 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 αangle
(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 tvalue 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 precalculated 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 precalculated, 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 xray 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 precalculated 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 3Dtexture 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 zcoordinate 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 precalculated 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
2π
....
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 tdependent 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
θdependent constant memory tables.
The main loop reconstructs one 2Dimage at a time. For each image, the corre
sponding zvalue 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 Qweight
9 WFBP: accumulate Qweight
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 Qweight
18 WFBP: accumulate Qweight
19 end
20 end
21 WFBP: Divide the pixel value with the accumulated Qweight
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 suboptimal. 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 conebeam CT: the RIWFBP
method based on iterative application of the WFBP method, the OSIWFBP, and
the WLSIFBP 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 nonlinear 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 zdirection. Fortunately, these can be avoided
by preprocessing of projection data with a lowpass ﬁlter in the qdirection.
To reduce the number of required iterations, the ordered subsets technique [39]
can be applied to RIWFBP, resulting in the OSIWFBP 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 OSIWFBP is
nonregularized, and its convergence properties are uncertain.
Due to the nonsymmetric 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 WLSIFBP method, which is similar to
the RIWFBP, but has a simpliﬁed normalization procedure. Thanks to the latter
feature, WLSIFBP 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 WLSIFBP 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 WLSIFBP methods. Experiments with WLSIFBP show that the
spatial resolution for high contrast structures can be improved while preserving the
spatial resolution and signaltonoise 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 DCaliasing
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 DCaliasing
in both forward projection and backprojection, and therefore reduces aliasing ar
tifacts in comparison with lineintegration 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 stripintegration 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 WLSIFBP. As shown in Chapter 7, the WLSIFBP
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 WLSIFBP could be improved by replacing the current
Qweighting 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 conebeam geometry. Since rebinning to semiparallel data is the ﬁrst
step in the methods studied here, inplane resolution is inevitably lost. By formu
lating the WLSIFBP method directly in the conebeam geometry, and thereby
entirely avoid rebinning, there is hope to improve spatial resolution and suppress
inplane 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
FieldProgrammable 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 OSIWFBP p. 6
Point Spread Function PSF p. 11
qGeneralized 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 WLSIFBP p. 7
B Notation
Below follows a list of symbols and acronyms used in the thesis. Unfortunately
the symbols α and β mean both fanbeam 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
Parallelbeam projection data coordinates θ, t p. 9
Fanbeam 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
Sourceisocenter distance R
S
p. 2
Detector height h p. 3
Table feed per turn/helical pitch P p. 3
Maximal fanangle β
max
p. 36
Maximal coneangle 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, nonrebinned 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 downweighting parameter Q p. 17
Regularization parameter β p. 34
Regularization parameter xyplane β
xy
p. 48
Regularization parameter zdirection 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 xray
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 conebeam 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 CTImages Using Correlation Analysis. IEEE Transactions on
Medical Imaging, 27(12):1685–1703, 2008.
[5] R. N. Bracewell. The Fourier transform and its applications. McGrawHill,
Singapore, 1986.
[6] R.N. Bracewell and A.C. Riddle. Inversion of fanbeam 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
3Dreconstruction for helical conebeam scanning of long objects. A new
detector arrangement and a new completeness condition. In 4th International
Meeting on Fully ThreeDimensional 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 PImethod for helical conebeam CT. Technical Report
LiTHISYR2085, ISSN 14003902, Department of Electrical Engineering,
Link¨ oping University, December 1998.
[12] P.E. Danielsson and M. Magnusson. A proposal for combining FBP and
ART in CTreconstruction. 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 2Dcase.
Technical Report LiTHISYR2634, ISSN 14003902, 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 ThreeDimensional
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 CTreconstruction. 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. Distancedriven 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 Xray 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 Xray 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
maximumlikelihood 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 parallelbeam 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 Xray 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. Conjugategradient preconditioning methods
for shiftvariant 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 16slice 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 zﬂying 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.unierlangen.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 threedimensional 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 threedimensional electron microscopy and xray photog
raphy. Journal of Theoretical Biology, 29(3):471–481, 1970.
[32] P.C. Hansen. Rankdeﬁcient and discrete illposed problems: numerical as
pects of linear inversion. Society for Industrial and Applied Mathematics,
1997.
[33] K.M. Hanson and G.W Wecksung. Local basisfunction 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 zresolution reconstruction. In 8th
International Meeting on Fully ThreeDimensional 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 singleslice re
binning in conebeam 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 backprojectiontype 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. SNRweighted 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 trilinear 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 conebeam CT. Physics in Medicine and Biology, 50(4):613–
623, 2005.
[53] J. S. Kole and F. J. Beekman. Evaluation of accelerated iterative xray
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. Preprocessing 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.unierlangen.de/phantoms.
[60] R. M. Lewitt. Multidimensional digital image representations using gen
eralized KaiserBessel 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 tradeoﬀ 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 3D conebeam reconstruction with the
simultaneous algebraic reconstruction technique (SART) using 2D texture
mapping hardware. IEEE Transactions on Medical Imaging, 19(12):1227–
1237, 2000.
[68] K. Mueller, R. Yagel, and J. J. Wheller. Antialiased threedimensional cone
beam reconstruction of lowcontrast 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 threedimensional reconstruction from conebeam 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 reprojection 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 CTscanner. In 9th International Meeting on Fully ThreeDimensional
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 nPI
method for helical conebeam 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. Threedimensional 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: nonsymmetrical 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 128Core 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 FIRFiltern 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 GPUBased 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 ConeBeam 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 conebeam reconstruction. Medical Physics, 31(8):2230–
2236, 2004.
[89] L. A. Shepp and B. F. Logan. Reconstructing interior head tissue from xray
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 multislice 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 conebeam com
puted tomography by constrained, totalvariation 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 multislice 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 maximumlikelihood 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.unierlangen.de/phantoms.
[98] S. Srivastava. Accelerated statistical image reconstruction algorithms and
simpliﬁed cost functions for xray computed tomography. Phd thesis, Univer
sity of Michigan, Ann Arbour, June 2008.
[99] S. Steckmann, M. Knaup, and M. Kachelriess. Hyperfast generalpurpose
conebeam spiral backprojection with voxelspeciﬁc 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 nonexact reconstruction in cone
beam CT. Master’s thesis, Link¨opings Universitet, September 2004.
[102] J. Sunneg˚ardh. Combining analytical and iterative reconstruction in helical
conebeam CT. Licentiate Thesis, Link¨opings Universitet, February 2007.
[103] J. Sunneg˚ardh and P.E. Danielsson. Iterative improvement of nonexact re
construction in conebeam 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 antialiased 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 conebeam CT. In 9th International Meeting on
Fully ThreeDimensional 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 conebeam CT. Medical Physics, 35(9):4173–4185,
2008.
[107] J. Sunneg˚ardh and M. Grasruck. Nonlinear regularization of iterative
weighted ﬁltered backprojection for helical conebeam CT. In IEEE Nu
clear Science Symposium and Medical Imaging Conference Record, pages
5090–5095, 2008.
[108] K. C. Tam. Threedimensional 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 threedimensional
statistical approach to improved image quality for multislice helical CT.
Medical Physics, 34(11):4526–4544, 2007.
[110] H. Turbell. ConeBeam 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 conebeam reconstruc
tion with standard and nonstandard 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. Ultrafast 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. Upsampling 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. Modelbased image reconstruction in Xray 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 errorreductionbased algo
rithm for conebeam computed tomography. Medical Physics, 31(12):3206–
3212, 2004.
[124] A. Ziegler, T. K¨ ohler, T. Nielsen, and R. Proksa. Iterative conebeam CT
image reconstruction with spherically symmetric basis functions. In 8th
International Meeting on Fully ThreeDimensional 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 PIlines from minimum
data in helical conebeam 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 conebeam 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 SE581 83 Link¨ping, Sweden o Copyright c 2009 Johan Sunneg˚ ardh Sunneg˚ ardh, Johan Iterative Filtered Backprojection Methods for Helical ConeBeam CT ISBN 9789173935869 ISSN 03457524
A Typeset with L TEX Printed in Sweden by LiUtryck, Link¨ping, 2009 o
iii
Abstract
Stateoftheart reconstruction algorithms for medical helical conebeam Computed Tomography (CT) are of type nonexact Filtered Backprojection (FBP). They are attractive because of their simplicity and low computational cost, but they produce suboptimal 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 nonexact 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 nonexactness 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 (WLSIFBP). 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 edgepreserving nonlinear regularization. In this case, resolution becomes contrast dependent, a fact that can be utilized for improving high contrast resolution without degrading the signaltonoise 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 nonlinear regularization on the other hand, the degraded signaltonoise 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 ickelinj¨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 ickeexakta 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 ickelinj¨r regularisering unders¨kts. f¨rkortad WLSIFBP. 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 ickeexakta. 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.. 22092218. 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 OSIWFBP. 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/brusf¨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 RIWFBPmetoden. 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 ickelinj¨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 WLSIFBP 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 proofreading. 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 proofreading. 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 proofreading 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. PerErik Danielsson who introduced me to the ﬁeld. and for making me happy. CMIV. Herbert Bruder. and for proofreading 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 Twodimensional 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 (OSIWFBP) 6. . 7. .2 Weighted least squares IFBP (WLSIFBP) . . . . . . . . . . . 10. . . . . . 8. . . . . . . . . . . . . .3 OSIWFBP . . . 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 nondestructive 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 Xray images acquired in multiple directions to calculate crosssectional images of the Xray 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 zaxis. High acquisition speed is generally desirable since this reduces examination time and image artifacts due to motion. which originates from 1972. consisting of an Xray 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 coneangle 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 zaxis.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 Xray source
x Figure 1.2: Scanning geometry for helical conebeam 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 nonexact. 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 nonexact 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, nonexact methods are computationally less demanding, simpler to implement, and oﬀer a better dose utilization than exact methods. One stateoftheart analytical nonexact 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). conebeam 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. Windmill artifacts are caused by aliasing in the zdirection of the detector, and look like alternating dark and bright streaks extending radially from the upper and lower ends of the spheres. The coneartifacts 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 nonexact 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 nonexact reconstruction in helical conebeam 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 wellknown 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 wellknown 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 OSIWFBP 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 (OSIWFBP).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 WLSIFBP can be used to improve high contrast spatial resolution and reduce artifacts while increasing the signal to noise ratio for lowcontrast 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 WLSIFBP produces images very similar to those of RIWFBP.5 Contributions and publications 7 Weighted Least Squares IFBP (WLSIFBP) 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 WLSIFBP and the qGeneralized Gaussian Markov Random Field (qGGMRF) 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 WLSIFBP 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 WLSIFBP.. 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. nonquadratic potential functions.
W . • Scalars and scalarvalued 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 WLSIFBP 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 xray 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 twodimensional 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 twodimensional 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 Twodimensional image reconstruction The mathematical problem of twodimensional 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 twodimensional Fourier transform Ff of a function f (x.3) The theorem is illustrated in Fig.1: Photon intensities of attenuated and nonattenuated xray beams ∫ diﬀer by a factor e − Γt. Performing a ﬁnal inverse twodimensional Fourier transform yields the reconstructed result. Theorem 1 can be applied directly to parallel projection data by computing the twodimensional Fourier transform of f (x. ρ sin φ) = (Ft (Rf ))(ρ. φ = θ − π . θ). However. Figure 2. 2. y) through a series of onedimensional 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 onedimensional 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 Twodimensional 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)ρei2πρ(x sin θ−y cos θ) dρdθ = 0 π −1 (Ft (Ft p · Fg∞ ))(x sin θ − y cos θ. this image will be nonzero 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 = ρdρdφ) 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 onedimensional 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) ρdρdφ =(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 lowpass 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 nonzero.. [100]. Therefore. i. Nt −1 q(t. (2. a continuous projection q(t. tNt −1 .e. is an approximative reconstruction method for helical conebeam 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 conebeam geometry. Right: the semiparallel 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 conebeam data to semiparallel data (Fig. 2.5), 2) Rampﬁltering 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 xyplane.
Figure 2.7: Scanning geometry in the yzplane. 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 xyplane, 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 xyplane, the rebinned projections look perfectly parallel. However, the rays diverge in the zdirection and the rebinned projections are therefore called semiparallel 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 lowpass 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 twodimensional 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 nonzero contributions in the sum over k.5. give rise to severe nonlocal 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 xray 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 nonexact 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 conebeam 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 SignaltoNoise 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 conebeam CT [73, 16, 121, 127, 109, 98]. • Incomplete data. One example of how to suppress artifacts due to missing data in circular conebeam 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 xrays 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 nonexactness 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 twodimensional array of image coeﬃcients are mapped onto a continuous function by interpolation with a bilinear basis function. The sampling distance in the ydirection is assumed to be equal to the sampling distance in the xdirection.
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 threedimensional 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 threedimensional 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 threedimensional 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 threedimensional 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 costeﬀective 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 nonzero 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 DCcomponents was named “DCaliasing” 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 DCcomponent will appear in the result.3. One wellknown example of rotationally symmetric basis functions satisfying (i) and (ii) above is the class of generalized KaiserBessel 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 xray 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 zdirection after semiparallel 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 antialiasing 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 DCaliasing. 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 noncentral repetitions of the DCcomponent 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 wellbehaved 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 KaiserBessel basis functions (also known as “blobs”) for iterative reconstruction. 3. This is caused by (i) illposedness 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 wellbehaved functions. A presentation and discussion of the general problem of illposedness 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. illposedness 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 illposed 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 suboptimal 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 zdirections 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 nonattenuated 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 Noisefree 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 submeasurements.
8 Active detector area (xyplane) dβ 0. .77◦ a Pitch angle γP 1.9 Active detector area (zdir) dz 1. both source and detector are divided into subunits. NDH 3 Source subsampling along xyplane NSW Source subsampling in zdir 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 IsocenterDetector 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 subcontributions.04◦ a This is the cone angle of a detector covering exactly the TamDanielsson window [110]. SourceIsocenter distance RS 595mm 490.1: Scanning parameters for experiments in following chapters.6mm Tablefeed 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 semiparallel 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 zdirection. 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 semiparallel 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 noisefree 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 inplane (xyplane) 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 stepbystep 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 edgebased 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 crossplane (zdirection) 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 inplane MTF is to use a thin wire test object directed orthogonal to the xyplane [46]. a thin plate orthogonal to the zaxis is often used to measure the crossplane 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 inplane ˜ .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 crossplane 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 nonlinear 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 abovementioned 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 ﬁnitedimensional 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 xray attenuation measurements. The iterative method.2). (1) Rebinning to semiparallel geometry (pin → preb ). The basic idea is to start with a nonexact 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 nonexact 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 preprocessed 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 lowpass 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 inplane and crossplane 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 lefthand side matrices corresponds to the convolution operations on the righthand 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. nonlocal 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 nonexact. 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 noniterative 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 highfrequent artifacts commonly known as spiral or windmill artifacts. g.5. the lowfrequency 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 twodimensional 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 closeups 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. Preﬁltering 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 zdirection. 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 preﬁltered 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 zaxis. 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 isocenter and 140mm away from the isocenter.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 oﬀcenter 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 dashdotted 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 nonregularized 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 TamDanielsson 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 PImethod [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 PImethod.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 nonexact 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 preprocessing suppresses artifacts nonlocal 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 nonstatistical 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 nPI detectors [80] (see Fig. [26]). However. A number of other approaches exist for alleviating windmill artifacts in helical conebeam 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. nonlinear 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: PIdetector. alleviating windmill artifacts. In this chapter. and other types of highpass ﬁlters. the PI detector and the 3PI 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 nonlinear 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 windmill 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: 3PI detector. resulting in a considerable alleviation of windmill artifacts. It should be mentioned that besides the above mentioned software approaches for windmill artifact reduction. By rapidly moving the focal spot between two diﬀerent zpositions. 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. nonlinear regularization has shown to be useful for preserving highcontrast 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 nonexistent. 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 columnaction 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 OSIWFBP method.8◦ ).Chapter 6 Ordered Subsets IWFBP (OSIWFBP) 6.1 Introduction This chapter is an adapted and reduced version of [102. presenting the OSIWFBP 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 conebeam 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 OSIWFBP eﬃciently suppresses cone artifacts for high cone angles. mainly because the OSIWFBP 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 nonregularized 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 OSIFBP 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 (OSIWFBP) 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 θoﬀset ({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 OSIWFBP 73 6.. the . certain noiselike 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 DCcomponent 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 twodimensional experiments with the Forbild thorax phantom [97] shown in Fig. as we will see in the next section. m = − M2 . The kernel used for lowpass 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 threedimensional 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)). rowwise lowpass 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 subiterations. The iteration of OSIWFBP is given by fk+1 = fk + αk Qik (preb. However. is chosen.3 OSIWFBP 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 OSIWFBP method is rebinning of the input data pin to semiparallel 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 (OSIWFBP) 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. SourceIsocenter distance Number of channels (Quarter oﬀset) Number of rows Number of projections per turn Detector height Tablefeed 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 σvalues determining the amount of lowpass 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 OSIWFBP 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 128row 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 64row 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 OSIWFBP. 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 (OSIWFBP) RIWFBP. α=0. and OSIWFBP 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 lowpass 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 lowpass ﬁlter in (6.25.4).6: Blown up twodimensional reconstructions of the Forbild Thorax phantom [97]. β = 0 b) OSIWFBP. showing highfrequency “ringing” artifacts. Greyscale window ±50HU. No LP c) OSIWFBP.78 Ordered Subsets IWFBP (OSIWFBP) a) RIWFBP. α = 0. . The number of subsets is ten.0. No LP d) OSIWFBP. 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 OSIWFBP 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. OSIWFBP 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 OSIWFBP 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 OSIWFBP. 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 OSIWFBP 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) OSIWFBP. β = 0. For the OSIWFBP reconstructions. f0 (b) OSIWFBP. and for e) and f).7: Clock phantom reconstructions with RIWFBP (β = 0) and OSIWFBP with 10 subsets.80 Ordered Subsets IWFBP (OSIWFBP) (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) OSIWFBP. α = 0. f20 Figure 6. Cone angle κ ≈ ±4.
f2 (f) OSIWFBP. β = 0.6◦ . f10 (e) RIWFBP. f1 (d) OSIWFBP. . 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) OSIWFBP.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 (OSIWFBP) 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 OSIWFBP 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 matrixvector 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 WLSIFBP. 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 nonnegative. . 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 (WLSIFBP) 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 nonnegative 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 (WLSIFBP) 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 nonnegative. (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 (WLSIFBP) 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 WLSIFBP 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 lowfrequency 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 nonnegative. 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 WLSIFBP 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 noniterative 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 zdirection of the reconstruction volume. two types of artifacts are present: (i) lowfrequency cone artifacts caused by the nonexactness of the reconstruction method. The Joseph backprojector PT does not satisfy this criterion.7.1: WLSIFBP 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 WLSIFBP using noisefree 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 WLSIFBP methods. βxy = 1. 7. Greyscale window ±20HU. βz = 1.3.5.3 Experiments 87 a) WLSIFBP. 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 DCaliasing [12. To avoid these artifacts. f5 Backprojector: PT b) WLSIFBP. 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 noniterative WFBP results in strong nonlocal artifacts. MTF measurements for the RIWFBP and the WLSIFBP 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 xyplane 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 WLSIFBP 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 WLSIFBP method. After ﬁve iterations. Fig. 7.8 shows noise measurements on clock phantom reconstructions for RIWFBP and WLSIFBP at approximately equal spatial resolution. using edge phantoms of radius 20mm placed in the isocenter. Fig. The MTFs for RIWFBP and WLSIFBP 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 WLSIFBP method than for the RIWFBP method. For both methods. As expected. there are only small diﬀerences between the RIWFBP and WLSIFBP reconstructions.5.7 was used for initialization. the RIWFBP and WLSIFBP results look very similar. However. and WLSIFBP 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) WLSIFBP.2: RIWFBP and WLSIFBP reconstructions of the clock phantom. f1 d) WLSIFBP. β = 0.
the similarity in formulation between WLSIFBP 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) WLSIFBP. Greyscale window ±20HU.7. 7. βz = 1. a small modiﬁcation of the method results in the WLSIFBP 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) WLSIFBP. 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 WLSIFBP 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 WLSIFBP does not .
Regularization parameters: β = 0. Q = 0.0. f5 . βxy = 1.0 Figure 7. RIWFBP.7.0 e) WLSIFBP. Q = 0. βz = 1.7 and Q = 1. Q = 1.7 f) WLSIFBP.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 WLSIFBP 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 WLSIFBP.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 WLSIFBP curves overlap both in the isocenter and oﬀcenter cases. 1 0.Q=0. f 5 . Q=1.2 and 7. f 5 .7 0.7 WLSIFBP. Q=0.3 50.1 50 49.3 0. r = 150mm WLSIFBP. 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 WLSIFBP method in the isocenter and 150mm oﬀcenter. r = 0mm WLSIFBP.0 WLSIFBP.3 50.
8 0.5 15 14.7 RIWFBP.3 WLSIFBP. Q = 0.2 0.9 0.5 17 16.8: σn measurements made on clock phantom reconstructions. r = 150mm WLSIFBP. f 5 . f 5 . . r = 0mm WLSIFBP.7 Normalized SSP 0.7: Normalized SSP measurements of the RIWFBP and the WLSIFBP method in the isocenter and 150mm oﬀcenter.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 WLSIFBP.7 WLSIFBP.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 nonlinearly regularized versions of the RIWFBP and WLSIFBP methods. employed to deﬁne the nonlinearity of the regularization operator. This property simpliﬁes the analysis and makes the results easy to predict: ignoring nonlinearities 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 WLSIFBP 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 WLSIFBP. 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 /cp−q (8. the qGGMRF functions are shown for diﬀerent values of q. The qGGMRF inﬂuence function is given by V (f ) = f p−1 1 + f /cp−q p− p−q cp−q f p−q 1 + f /cp−q sgn(f ). The function V (f ) is called potential function. the value of c determines where the intersection between the qGGMRF 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 nonconvex and convex potential functions have been suggested for image reconstruction and restoration. [21]) and the qgeneralized Gaussian Markov Random ﬁeld (qGGMRF) 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 nonconvex potential functions are the saturated quadratic function and the Geman and McClure potential function [29].4) The parameters p. we will focus on the qGGMRF function V (f ) = f p . (8. the potential function becomes increasingly vshaped. 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: qGGMRF potential functions for diﬀerent values of q and c. q = 1.2: qGGMRF 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 nonlinear regularization we limit the experiments in this chapter to the WLSIFBP 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 preprocessed as little as possible. more iterations are required for convergence. and 500HU. 80HU. which is the traditional choice. Some crossplane resolution contributions seem to remain. the new update formula becomes fk+1 = fk − αS(Q(Pfk − preb ) + βRV fk ). projection data have been preﬁltered as described in Chapter 5. the results of WLSIFBP 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 inplane 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 inplane 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 crossplane 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 nonregularized 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 nonregularized 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 nonnormalized 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 WLSIFBP(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 WLSIFBP in Fig. Here. The image is corrupted by streak artifacts plus a combination of coneartifacts. To the left. Fig. attained with nonlinear regularization (p = 2. the linear WLSIFBP method gives a better result with less pronounced windmill artifacts. To examine the noise properties for the nonlinearly regularized WLSIFBP. 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 WLSIFBP 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 WLSIFBP using quadratic regularization. σnq are noise measurements for the quadratically regularized WLSIFBP method. we have studied head phantom reconstructions for diﬀerent numbers of nonattenuated photons. i.. In relation to the linearly regularized WLSIFBP method.3. For any linear reconstruction method. for nonlinear methods. as well as improves the crossplane 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. windmill artifacts.10 b) shows an axial image of a WFBP head phantom reconstruction from noisefree data. and nonlinear WLSIFBP. 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 WLSIFBP. 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) WLSIFBP.3.10: Axial images of noisefree head phantom reconstructions. f10 . f10 . Greyscale window:−25HU. q = 1. . c = 20 Figure 8.25HU.0. p = 2. p = q = 2 d) WLSIFBP.
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: WLSIFBP 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 WLSIFBP 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) WLSIFBP. f10 .3 Experiments 107 a) WFBP reconstruction b) WLSIFBP. 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 windmill 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 onedimensional 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 lowpass 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 WLSIFBP 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 WLSIFBP reconstruction with respect to windmill artifact reduction and high contrast spatial resolution.1. For the nonlinear WLSIFBP method. by increasing β to 0. 8. 8.6 does not improve the result much over the linear WLSIFBP. 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 WLSIFBP 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 suboptimums 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) WLSIFBP.3.8. Greyscale window: ±150HU.0 c) WLSIFBP. q = 1. c = 50. f10 . c = 50.4 Discussions and conclusions 109 a) WFBP reconstruction b) WLSIFBP.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 qGGMRF potential function adds a number of regularization parameters to the cost function.15: Onedimensional 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 WLSIFBP 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 viewdependent 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 viewindependent trilinear 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 conebeam 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 conebeam 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 zaxis as shown in Fig. (9. and focus on improving the reconstruction by modeling the lowpass 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 xyplane. 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 zaxis. RS sin α. we accept the lowpass 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 xyplane.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 xzplane. 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. DCaliasing 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 isocenter and one located 150mm oﬀcenter for measuring oﬀcenter 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. DCaliasing 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 subangles used to simulate this eﬀect.
118 Data acquisition modeling Figure 9. 9. These DCaliasing artifacts occur mainly because the interpolation function in the zdirection 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 WLSIFBP 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 DCaliasing artifacts is reduced as shown in Fig. NDW . Fig. and NGR have negligible impact on the crossplane resolution.2mm.5 c). We begin with studying crossplane resolution in the isocenter. It is reasonable to assume that the other parameters NSW .4: Illustration of DCaliasing during backprojection.5 d). 9. modeling of the detector element height with NDH = 3 will always be used for suppressing DCaliasing 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) WLSIFBP. PT NSL = 1. NDW = 2 Figure 9. NDW = 2 f) WLSIFBP. f10 . NDW = 1 d) WLSIFBP. 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) WLSIFBP.5: Axial images of MTF phantom reconstructions for diﬀerent backprojectors and number of rays per measurement. NDW = 1 c) WLSIFBP. 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 crossplane resolution case. these parameters have all been set to one.25 0 5 10 Iteration number Data acquisition modeling BS . consider . Inplane resolution can be divided into radial resolution in the isocenter. 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) oﬀcenter 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 oﬀcenter 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 crossplane 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 WLSIFBP.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: Blowup 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 WLSIFBP 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: isocenter MTF ρ50% values for WLSIFBP.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 oﬀcenter radial resolution.9: Left: oﬀcenter radial MTF ρ50% values for WLSIFBP.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 oﬀcenter. 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 fanangle 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 xray 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 oﬀcenter 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 oﬀcenter 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 nonlinear regularization. N GR = 2 PT . Fig. in crossplane resolution at a reasonable cost increase. Unfortunately. The PT result has a slightly higher crossplane resolution.6 0.10: Left: oﬀcenter azimuthal MTF ρ50% values for WLSIFBP. 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 crossplane resolution is further improved by the detector element modeling.12 shows how nonlinear regularization can partially be used to control the increased amount of windmill artifacts.8 MTF. but as indicated by the dashed circles. We note that the detector element modeling improves the crossplane 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 windmill 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 windmill 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: WLSIFBP 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 crossplane 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 inplane resolution than PT . Δq h/2 = 1. Greyscale window (40 ± 75)HU. the inplane 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 WLSIFBP 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, inplane
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, inplane
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, crossplane
d) Fourier domain, crossplane
Figure 9.13: Interpolation functions and their Fourier transforms for BS and PT in the tdirection and qdirection respectively. BS than for PT . Another reason is that the balance between the regularization contribution and data contribution is aﬀected by the amount of lowpass 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 oﬀcenter 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 inplane resolution is limited by initial rebinning of input data. One possibility to further improve inplane resolution could be to entirely avoid rebinning by performing forward projection and backprojection directly in the conebeam geometry, and use a rampﬁlter derived for fanbeam 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 inplane 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 DCaliasing 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 2D parallel beam CT. Except for some small details, generalization to 3D conebeam 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 DCaliasing 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 Pmatrix 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 Pmatrix 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 WLSIFBP 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 subrays. t1 − xj ). Noisefree input data were generated from a SheppLogan 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 .. 2D 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 SheppLogan 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 DCaliasing 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 semiparallel 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 parallelbeam 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 SheppLogan phantom after ten iterations. while σe2 measures errors on low contrast regions.
σe1 measures errors over the whole image.17: Reconstructed images of the SheppLogan 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 conebeam 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 precalculated 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 conebeam geometry where the adaptive antialiasing 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 lookup 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 zdependent angle φ = φ0 + 2πz/P . since all coeﬃcients are precalculated. 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 nonzero 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 wellknown 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 lookup 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 FieldProgrammable 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 3D 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 readonly 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. 2D. 10. The texture cache improves memory access from 1D. 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 400600 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 1D. the global memory . and simple and fast handling of boundary conditions. All arrays are assumed to have the length N. or 3D 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. 2D. 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 forloop. 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 ifstatement 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.ifstatement 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 2D 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 1D. 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. 2D. 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 3D 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 tvalue 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 precalculated and made accessible to the kernel through . For each ray/thread. Calculate constant memory lookup tables for simpliﬁed computation of the xray source position (α. the input volume is stored as a threedimensional cudaArray. the αangle (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 3Dtexture 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 precalculated. the task of one instance of this program is to calculate one line integral. Accumulate the contribution. Calculate source position with help from the precalculated 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 xray 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 zcoordinate 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 precalculated 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 3Dtexture 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 tdependent 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 2Dimage at a time. The ﬁrst step in the kernel is to calculate the voxel indices xIx and yIx from blockIdx and threadIdx. the corresponding zvalue 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 θdependent 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 Qweight WFBP: accumulate Qweight end end WFBP: Divide the pixel value with the accumulated Qweight 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 Qweight WFBP: accumulate Qweight 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 suboptimal. 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 WLSIFBP 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 nonlinear 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 zdirection. the OSIWFBP. 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 conebeam 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 lowpass ﬁlter in the qdirection. 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 OSIWFBP method.
and therefore reduces aliasing artifacts in comparison with lineintegration 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 nonsymmetric QP matrix. we present the WLSIFBP 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 WLSIFBP 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 DCaliasing during backprojection with PT . in its current form. Therefore. and hence diﬃcult to analyze. However. the OSIWFBP is nonregularized. Furthermore. WLSIFBP 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 DCaliasing 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 WLSIFBP show that the spatial resolution for high contrast structures can be improved while preserving the spatial resolution and signaltonoise ratio for low contrast structures. certain artifacts are slightly more pronounced for WLSIFBP 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 WLSIFBP. Unfortunately. the WLSIFBP 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 WLSIFBP could be improved by replacing the current Qweighting 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 stripintegration 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 inplane 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. inplane 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 semiparallel 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 conebeam geometry. By formulating the WLSIFBP method directly in the conebeam 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 FieldProgrammable 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 qGeneralized 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 OSIWFBP PSF qGGMRF ROI RIWFBP RMSE SP SNR SART SIRT SPECT SSP SM WFBP WLSIFBP 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 Parallelbeam projection data coordinates Fanbeam projection data coordinates Projection data row coordinate Frequency coordinates Radial frequency coordinate Sourceisocenter distance Detector height Table feed per turn/helical pitch Maximal fanangle Maximal coneangle 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 fanbeam 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. nonrebinned 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 downweighting parameter Regularization parameter Regularization parameter xyplane Regularization parameter zdirection 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. McGrawHill. 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 ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. Flohr. 3D ﬁltered backprojection for curved detector axial conebeam 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 fanbeam 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 3Dreconstruction for helical conebeam scanning of long objects. T. Kamphuis. 2007. N. Leng. J. Wavelet Based Noise Reduction in CTImages 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 xray 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 CTreconstruction. In IEEE Nuclear Science Symposium Conference Record. In 8th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. S. Utah. J. G. pages 2708–2710. De Man and S. Marchal. Dupont. Suetens. The 2Dcase. P. [15] P. Nuyts. ISSN 14003902. 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 Xray 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 LiTHISYR2085. Sunneg˚ ardh. Basu. ISSN 14003902. J. and H. Magnusson. 5(5):740–753. C. 20(10):999–1008.E. Advanced linear modeling and interpolation in CTreconstruction. 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 LiTHISYR2634. 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 parallelbeam tomography. Combining Fourier and iterative methods in computer tomography. Link¨ping University. Danielsson. Bresler. 2004. Danielsson and M. The original PImethod for helical conebeam CT. J. Sunneg˚ ardh. 2001. o [14] P. 47(3):977–981. An iterative maximumlikelihood 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 Xray 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. Distancedriven 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 zﬂying 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 threedimensional electron microscopy and xray 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.unierlangen. Herman. Flohr. Image reconstruction and image quality evaluation for a 16slice 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 basisfunction approach to computed tomography. Conjugategradient preconditioning methods for shiftvariant 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. Rankdeﬁcient and discrete illposed problems: numerical aspects of linear inversion. A. Image reconstruction from projections. Ulzheimer. McClure. and M. A. D. Iterative methods for the threedimensional 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 Xray 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 singleslice rebinning in conebeam 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 zresolution 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 ThreeDimensional 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 backprojectiontype 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. SNRweighted 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 trilinear 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 KaiserBessel window functions. Sweden. Lewitt. M. Preprocessing 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.unierlangen. 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 xray CT image reconstruction using ﬂoating point graphics hardware. 1984. Evaluation of the ordered subset convex algorithm for conebeam 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 tradeoﬀ 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. Antialiased threedimensional conebeam reconstruction of lowcontrast 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 ThreeDimensional 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 reprojection 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 3D conebeam reconstruction with the simultaneous algebraic reconstruction technique (SART) using 2D texture mapping hardware. [68] K. Fast implementations of algebraic methods for threedimensional reconstruction from conebeam 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 CTscanner. 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 FIRFiltern 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 multislice 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 ConeBeam CT on the Cell Broadband Engine Architecture. Scherl. edia tors. Vienna. 2000. The frequency split method for helical conebeam reconstruction. Wahl. Kowarschik. Birkh¨user Verlag. and J. The nPImethod for helical conebeam 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 xray 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 GPUBased CT Reconstruction using the Common Uniﬁed Device Architecture (CUDA). Ryoo. Program Optimization Study on a 128Core 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. Threedimensional reconstruction from radiographs and electron micrographs: application of convolutions instead of Fourier transforms. The approximate inverse and conjugate gradient: nonsymmetrical 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.unierlangen. volume 5032 III. Link¨pings Universitet. 2008. Germany. [98] S. Thomas Jr. Noise and edge artifacts in maximumlikelihood reconstructions for emission tomography. June 2008.E.. Sunneg˚ ardh. Iterative improvement of nonexact reconstruction in conebeam CT. Sunneg˚ ardh and P. Mori. A new antialiased 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 multislice 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. totalvariation 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 conebeam 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 generalpurpose conebeam spiral backprojection with voxelspeciﬁc weighting. and I. Steckmann. In IEEE Nuclear Science Symposium Conference Record. Danielsson. Image reconstruction in circular conebeam 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 xray 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 nonexact 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 conebeam reconstruction with standard and nonstandard 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. Ultrafast 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 conebeam 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 conebeam CT. In 9th International Meeting on Fully ThreeDimensional Image Reconstruction in Radiology and Nuclear Medicine. 2004. [112] E. ConeBeam reconstruction using ﬁltered backprojection. A threedimensional statistical approach to improved image quality for multislice helical CT. Threedimensional 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 conebeam 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 errorreductionbased algorithm for conebeam 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 conebeam 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 PIlines from minimum data in helical conebeam 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. Upsampling 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. Modelbased image reconstruction in Xray 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 ThreeDimensional 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 conebeam CT o image reconstruction with spherically symmetric basis functions.