Journal of Cardiovascular Computed Tomography (2011) 5, 286–292

Invited Review

New iterative reconstruction techniques for cardiovascular

computed tomography: How do they work, and what are the
advantages and disadvantages?
Rendon C. Nelson, MD, FACR*, Sebastian Feuerlein, MD, Daniel T. Boll, MD

Department of Radiology, Duke University Medical Center, Box 3808, Durham, NC 27710, USA

KEYWORDS: Abstract. The radiation doses associated with diagnostic CT scans has recently come under scrutiny. In
Computed tomography; the process of developing protocols with lower doses, it has become apparent that images reconstructed
Multidetector row CT; with a filtered back projection (FBP) technique are often inadequate. Although very fast and robust,
Helical CT; FBP images are prone to high noise, streak artifacts and poor low contrast detectability in low dose
Reconstruction situations. Manufacturers of CT equipment have responded to this limitation by developing new image
techniques; reconstruction techniques that derive more information from the data set. These techniques are based
Filtered back projection; on the use of maximum likelihood algorithms and are referred to at iterative reconstructions. This it-
Iterative reconstruction erative process can be used on the slice data alone, a combination of raw and slice data or on the raw
data alone. The latter approach, which is referred to as model based iterative reconstruction, is the most
computationally demanding as it models the entire process, from the shape of the focal spot on the
anode, the shape of the emerging x-ray beam, the three-dimensional interaction of the beam with
the voxel in the patient and the two-dimensional interation of the beam with the detector. This article
discusses the fundamentals of iterative reconstruction techniques, the pros and cons of the various man-
ufacturer approaches and specific applications, especially to cardiovascular CT.
Ó 2011 Society of Cardiovascular Computed Tomography. All rights reserved.

Introduction converge within the patient. Filtered back projection

Computed tomography (CT) originated in the mid-
1970s, and, since that time, essentially all cross-sectional
images have been made with the use of reconstruction
techniques that are based on the back projection of atten-
uation line integrals acquired from radial views that
Conflict of interest: R.C.N. is a consultant with GE Healthcare, Inc.
D.T.B. has received research support from Siemens Healthcare, Inc, and
Philips Healthcare, Inc. S.F. reports no conflicts of interest.
* Corresponding author.
E-mail address: rendon.nelson@dm.duke.edu
Submitted December 17, 2010. Accepted for publication July 13, 2011.
(FBP) algorithms are currently the most widespread
implementation of this technique.1–6 The main advantages
of the FBP include reconstruction speeds of %30–40 slices
per second, overall robustness, and relatively acceptable
image quality for most cases. The main disadvantage of
the FBP is the overall limitations in image quality, mainly
because of the back-projection process and the necessary
filtering, which accentuates noise and can result in streak
artifacts.7–9 Clinical situations in which image quality tends
to be less acceptable include: (1) CT scans in large patients
(Fig. 1); (2) images with small voxels, such as the targeted
field-of-view used for cardiac CT; (3) images through the
bony pelvis; and (4) intentionally low-dose datasets. An

1934-5925/$ - see front matter Ó 2011 Society of Cardiovascular Computed Tomography. All rights reserved.
doi:10.1016/j.jcct.2011.07.001
Nelson et al New iterative reconstruction techniques for CCT
Figure 1 CT scan of the mid-abdomen obtained in the axial
plane in a large patient. Note the high level of noise and streak ar-
tifacts as a result of aliasing or undersampling, and an artifact on
the left side of the patient where there is tissue outside the scan
field-of-view (arrow).
alternative to the FBP is an iterative reconstruction (IR) that
uses a maximum likelihood (ML) algorithm. IR techniques
are now commonly available from commercial suppliers;
therefore, familiarity with these techniques is necessary.
In this review, the concepts of IR are discussed, deliberately
avoiding the mathematical detail. The objective is to pro-
vide an intuitive ‘‘understanding’’ of the topic, which may
aid any person who is expected to use IR in practice.
Because of the advantages of using a FBP noted above,
there have been few incentives over the years for the
manufacturers to develop new reconstruction algorithms.
This fact began to change, however, with the introduction
and propagation of multidetector row CT scanners. With
this technology, several new applications have emerged,
including coronary CT angiography (CTA), pulmonary
CTA, CTA run-offs, CT colonography, and perfusion CT.
Perhaps more importantly, the speed and isotropic nature of
multidetector row CT scanners is well suited to the study of
patients with emergent (eg, trauma) or urgent (eg, acute
abdominal pain) conditions. As a result, the number of CT
scans currently performed has dramatically increased since
1998 when the 4-detector row scanners were introduced.
For several reasons, the radiation doses associated with
early multidetector row CT scans were higher than with
single detector CT scans.10 As a result, both the radiologic
and the nonradiologic communities began to take a strong
interest in the delivery and effects of ionizing radiation in
patients. Although there is particular concern in performing
CT scans on patients that have a higher sensitivity to ioniz-
ing radiation, specifically children and young adults, there
has been an overall concern about the administration of ion-
izing radiation to the general population, as well.
287
Filtered back projection
What exactly is a FBP and why is it deficient in low-
radiation dose CT examinations? Basically, a FBP is a
technique whereby projections or views from the raw
data are first filtered then back-projected to form a
2-dimensional array of voxels that are assigned a shade
of gray proportionate to the average x-ray photon dis-
tribution of each voxel.1–6 This reconstruction technique,
however, has significant limitations, because it relies on
several assumptions. These assumptions include the follow-
ing: (1) the focal spot comprises a single point at the
surface of the anode, (2) the x-ray beam is modeled as a
pencil-shaped ray passing through the object being
scanned, (3) there is a point-like interaction of the x-ray
beam with the voxel in the patient or object being scanned,
and (4) there is a point-like interaction of the x-ray beam
with the detector. In actuality, all four assumptions about
the x-ray beam are incorrect, thereby leading to substantial
limitations in spatial resolution and noise generation. Fur-
thermore, there is no statistical consideration for noise.
As a result, images acquired with a low-radiation dose tech-
nique and reconstructed with FBP tend to show a large
amount of noise, substantial streak artifacts, and poor
low-contrast detectability. With the current trend of reduc-
ing the radiation doses associated with CT in all patients,
regardless of radiation sensitivity, this reconstruction tech-
nique is no longer considered to be optimal.
Iterative reconstruction
What is an IR and why does it have advantages in low
radiation dose situations? First, let us define what is meant
by the term ‘‘iteration’’. An iteration is defined as a
‘‘problem-solving or computational method in which a
succession of approximations, each building on the one
preceding, is used to achieve a desired degree of
accuracy.’’11 This can be illustrated by a simple game in
which one person tries to guess another’s age on the basis
of intermediate feedback. Let’s say that you are trying to
guess the age of a 68-year-old stranger. If you assume the
person is between the ages of 0 and 100 years, the first
guess, splitting the difference would be 50. When you are in-
structed that your guess is too low, your next guess, again
splitting the difference would be 75. As this game pro-
gresses, you choose 63, then 69, then 66, finally arriving at
68 on your sixth guess.
Although there are many IR techniques, the description
here is restricted to ML reconstruction, the most commonly
used approach.12 ML reconstruction usually is performed
with the use of the expectation maximization (EM) algo-
rithm. The steps that occur during an IR are similar to those
taken during the game described earlier. (1) An initial esti-
mate of the x-ray photon distribution is made. (2) On the
basis of this starting point, an estimate is made of the
288 Journal of Cardiovascular Computed Tomography, Vol 5, No 5, September/October 2011
Figure 2 (A) Phantom having 2 white offset squares on a black
background. The arrows represent the projection angles at which
the dataset is sampled. (B) Sinogram of the 2-square phantom in
panel A, showing the sine wave that is obtained from a series of
projections through the phantom at various angles. Note that
this sinogram represents only one-half of the CT gantry rotation
(0 –180 ). The dotted lines correspond to the following angles
x-ray detector counts that would be acquired in each projec-
tion (with the use of forward projection) (Fig. 2). (3) The
estimated (forward-projected) projections are compared
with the actual measured projections acquired by the CT
system’s x-ray detector array (for EM the ratio of measured
to estimated x-ray detector counts is calculated). (4) This
comparison is then used to update the original estimate
(for EM this involves back-projecting the ratio and multi-
plying the original estimate by this correction image). (5)
The complete process is repeated, continuously estimating
the projections that would result from the revised x-ray
photon distribution.
Analogously to the number game, the iterative process
results in the estimated x-ray photon distribution getting
closer and closer to the true x-ray photon distribution. The
algorithm is said to converge to a solution, whereas the
changes between successive iterations are minimal. The
forward projection represents an inverse back projection;
however, unlike FBP, no filtering was necessary to reach a
solution.
Advantages and disadvantages of IR
Modeling the x-ray generation and detection process
One advantage of IR techniques is that they permit the
detection process to be accurately modeled. Differing from
FBP algorithms that make no allowance for the physics of
emission and detection of x-ray photons, the modeling
process can be quite comprehensive, including variations of
resolution with source-detector position, incorporation of
measured variable x-ray photon distributions, as well as
modeling of scatter in 3 spatial dimensions.
Noise properties
The eventually reconstructed noise patterns tend to be
much more acceptable than that present with the use of
FBP. In FBP, whereby noise is assumed to be constant
across the reconstructed field, noise patterns in iteratively
reconstructed datasets with the use of EM are correlated
with the signal; that is, the noise amplitude is lower in
regions of decreased level of x-ray distributions. The
absence of streak artifacts, which are commonly experi-
enced in noisy FBP studies, is a direct result of using IRs.
To maintain smoothness of the reconstruction, regulariza-
tion is introduced, which usually involves control of the
difference between neighboring pixels throughout the
reconstruction process. An alternative approach for noise
of projection: at 0 the projection sees 2 squares; at 45 the pro-
jection sees only 1 shape because 1 square hides behind the other;
at 90 the projection sees 2 squares; at 135 the projection sees
2 squares that appear wider than at other angles because they
are viewed from an oblique angle; and at 180 the projection
sees the 2 squares identical to the 0 angle.
Nelson et al New iterative reconstruction techniques for CCT 289

Figure 3 Depiction of the assumptions made relative to the

x-ray beam and the interaction with both the patient and the detec-
tor. Note that the ideal system assumes that there is a point source
focal spot on the anode, a pencil-shaped beam emerging from the
anode, a point-like interaction of the beam with the voxel, and a
point-like interaction of the beam with the detector. Note that in
actually none of these conditions exist.

reduction, from which similar results can be obtained, is to

use post-reconstruction filtering.

Accelerated EM with the use of ordered subsets

Various techniques can be used to improve the speed of

iterative reconstruction algorithms. A commonly applied
technique is called ordered subsets EM (OSEM). In previ-
ously described EM, the comparison is made between
estimated and measured projections for each iteration. This
requires considerable computational power because the
projections have to be calculated at typically multiple
angles, with each calculation of every projection taking at
least as long as a complete FBP reconstruction. With the
use of OSEM, the comparisons and subsequent updates are
based on only a limited number of projections from each

Figure 5 (A) FBP of the chest in the coronal plane obtained

Figure 4 Depiction of what characteristics of the x-ray beam are
modeled by the MBIR technique. This includes modeling of the
shape of the focal spot on the anode, the shape of the beam as
it emerges from the anode, the 3-dimensional interaction of the
beam with the voxel in the patient, and the 2-dimensional interac-
tion of the beam with the detector.
with an estimated radiation dose of 0.09 mSv. (B) ASiR of the
chest in the coronal plane with an estimated radiation dose of
0.09 mSv. (C) MBIR of the chest in the coronal plane with an es-
timated radiation dose of 0.09 mSv. Note superior depiction of the
blood vessels (white arrow) and bronchi (black arrow). Courtesy
of Universitair Ziekenhuis Brussels, Belgium.
290 Journal of Cardiovascular Computed Tomography, Vol 5, No 5, September/October 2011

iteration, progressively using varying projections in each
further iteration. With the use of OSEM, the algorithm will
target several subiterations before reaching its end point.
Because fewer calculations have to be done for each
subiteration, the computation is much faster.
Artifacts
ML reconstructions improve, in particular, reconstruc-
tion of noisy data, resulting in less evident streaking.
However, different and novel artifact patterns have been
introduced by IR techniques, particularly edge definition
artifacts in sharp transition zones in areas with differing
attenuation characteristics. These artifacts appear similarly
to the ringing artifacts commonly observed when using
restoration filtering. In general, artifacts tend to be less
evident with ML than with FBP.
Examples of currently available IR
techniques
IR of CT images can be performed from the image or
slice data, the projection or raw data, or both. Different
manufacturers of CT scanners use different methods to
reconstruct images with lower noise and in some cases
better low-contrast detectability and few artifacts. The main
objective for using these techniques is to achieve a
substantial reduction in the radiation dose either with the
same image quality or even superior image quality com-
pared with that which imagers have become accustomed to.
Several IR reconstruction techniques are currently avail-
able, and they can be categorized as: (1) IRs performed
from the image or slice data only, (2) IRs performed from
both the projection and image data, and (3) IR performed
from projection or raw data only.
IR from image data only
An example of an IR that uses the image or slice data
alone is IRIS (Iterative Reconstruction in Image Space;
Siemens Medical Solutions, Inc, Malvern, PA). With this
technique, the raw data are first reconstructed in the
traditional fashion with the use of a FBP. This information
is then forward-projected with multiple iterations according
to modeling of the noise data. The reconstruction time is
only slightly longer than that for the FBP. One of the
limitations, however, is that, because it uses the slice data
from the FBP, it continues to assume an ‘‘ideal system’’
(Fig. 3). Preliminary data suggest that IRIS significantly
diminishes the image noise and on the average can achieve
a 30%–40% dose reduction. Low-contrast detectability and
streak artifacts are not significantly improved.13 Further-
more, there is a slightly different appearance of these
IRIS images than FBP images, to which imagers tend to
adapt in a short period of time.

Figure 6 (A) Curved planar reformation through a coronary artery stent in the coronal and axial planes, reconstructed with a FBP tech-
nique. (B) Curved planar reformation through a coronary artery stent in the coronal and axial planes, reconstructed with MBIR. Note that
with MBIR the struts of the metallic stent are much sharper, and there is superior depiction of neointimal hyperplasia in the mid-portion of
the stent (arrows). Courtesy of Centre Cardiologique du Nord, France.
Nelson et al New iterative reconstruction techniques for CCT 291

IR from both the projection data and image data helical pitch, and whether a targeted field of view is speci-
fied. That is, datasets that have a large number of slices
Statistics-based IR (eg, a CT of the chest, abdomen, and pelvis) that use a helical
There are several different techniques that perform an IR pitch ,1.0, and/or have a small targeted field of view,
with both the image or slice data and the projection or raw
data. These include ASiR (Adaptive Statistical Iterative
Reconstruction; GE Healthcare, Inc, Waukesha, WI), Safire
(Sinogram Affirmed Iterative Reconstruction; Siemens
Healthcare Solutions, Inc), and iDose (Philips Healthcare,
Inc, Andover MA). With these techniques, the projection
data are first reconstructed with a FBP then compared with
an ideal noise model that is based on statistics, taking into
account both photon and electronic noise. Multiple itera-
tions are then performed that compare each updated voxel
with the ideal noise model until the algorithm con-
verges.14,15 The reconstruction times for these techniques
are slightly longer than that for the FBP. One of the limita-
tions, however, is that like IRIS these techniques continue
to assume an ideal system (Fig. 3). Furthermore, to improve
acceptability among imagers, the ASiR images are typi-
cally blended with FBP images with the use of a 40%–
50% weighting factor. To achieve a 50% blend, the data
from a voxel using the FBP are added to the data from
the same voxel using ASiR and then divided by 2. Prelim-
inary data, especially with ASiR which has been around
longer, suggest that there is a significant reduction in image
noise and that on the average, a 40%–50% dose reduction
can be achieved (Fig. 4).16–23 There may be a slight im-
provement in low-contrast detectability and slightly fewer
streak artifacts. Furthermore, there is a slightly different ap-
pearance of these images compared with FBP images, to
which imagers tend to adapt in a short period of time.

Full IR from the projection data only

Model-based IR (MBIR or Veo; GE Healthcare, Inc) is

the most advanced of the various IR schemes in that it
attempts to model the entire x-ray beam as it travels from the
cathode to the detector.24 That is, MBIR models: (1) the
2-dimensional shape of the x-ray beam on the anode as the
focal spot, (2) the actual shape of the x-ray beam as it leaves
the anode, (3) the 3-dimensional interaction of the x-ray
beam within the voxel in the patient, and (4) the
2-dimensional interaction of the x-ray beam with the detec-
tor array. By modeling these optical effects, MBIR actually
improves the spatial resolution of the image in all 3 dimen-
sions (Fig. 5). With the use of these various models, multiple
iterations are performed, and each iteration is compared with
the complex model until the algorithm converges. Because
MBIR is a complicated algorithm, using multiple iterations
and multiple models, the reconstruction time, even with ded-
icated state-of-the-art parallel processors, is significantly
longer than FBP, as well as the other IR techniques. Recon-
struction times range from 10 to 90 minutes and are deter-
mined by the number of slices to be reconstructed, the
Figure 7 (A) FBP image of the right shoulder in the coronal
plane reconstructed with a soft tissue algorithm. Note the horizon-
tal streak artifacts because of aliasing from low x-ray photon
counts through the bones. (B) FBP image of the right shoulder
in the coronal plane reconstructed with a bone algorithm. Note
the horizontal streak artifacts because of aliasing are even more
noticeable with this reconstruction. (C) MBIR image of the right
shoulder in the coronal plane depicted with a soft tissue window
and level. Note excellent depiction of the soft tissue about the
shoulder and significantly fewer streak artifacts. (D) MBIR image
of the right shoulder in the coronal plane depicted with a bone
window and level. Note excellent depiction of the comminuted
fracture of the coronoid process (arrow). Courtesy of the Univer-
sity of Wales, United Kingdom.
292 Journal of Cardiovascular Computed Tomography, Vol 5, No 5, September/October 2011
require more reconstruction time. With the use of parallel
processors, about 3–4 datasets can be reconstructed per
hour. Because of these longer reconstruction times, however,
initial application of this technique to clinical practice will
mainly focus on patients with nonurgent or nonemergent
conditions. Fortunately, in most practices the majority of
CT scans are performed in the outpatient setting, and imme-
diate assessment is not mandatory. Furthermore, it is possi-
ble to have a preliminary set of ASiR images for
immediate review. Preliminary data suggests that there is a
significant reduction in image nose and that radiation dose
reductions averaging 60%–70% can be achieved (Figs. 4,
6, and 7). Specific examples in which this technique might
be advantageous include ultralow dose CT of the chest as a
replacement for the chest radiograph (Fig. 4) and visualiza-
tion of the patency and integrity of vascular stents (Fig. 6).
Furthermore, because of the sophisticated modeling, there
is an improvement in spatial resolution and low-contrast de-
tectability with fewer streak artifacts. This is particularly ap-
parent when performing CT through the upper chest x-ray
due to absorption by the bones and soft tissues of both shoul-
ders (Fig. 7). One of the first impressions that imagers tend to
have on inspection of these images is that they have a differ-
ent ‘‘look and feel’’ compared with images reconstructed
with a FBP. The appearance has been described by some
as being a bit ‘‘waxy,’’ ‘‘plastic,’’ or ‘‘impressionistic.’’ How-
ever, comparison of FBP and MBIR images acquired with
the same radiation dose shows that, although there may be
a difference in the appearance of the images, the information
is essentially the same. As a result, imagers interpreting
these images tend to adapt to the new look in a relatively
short period of time, particularly if they have preliminary ex-
perience with images from other IR techniques.
Conclusions
IR algorithms provide several advantages to FBP, in-
cluding the ability to reduce noise in low-radiation dose
datasets. Some algorithms can even improve low-contrast
detectability while diminishing streak artifacts. The com-
bination of faster computers and acceleration algorithms
such as OSEM mean that the IR technique can be applied to
clinical practice.
References
1. Cormack AM: Representation of a function by its line integrals, with
some radiological applications. J Appl Phys. 1990;17:967–82.
2. Gordon R, Herman GT: Three-dimensional reconstruction from pro-
jections: a review of algorithms. In: Bourne GH, Danielli JF, editors.
International Review of Cytology. New York, NY: Academic Press,
1974. p. 111–151.
3. Herman GT, Lakshminarayanan AV, Rowland SW: The reconstruction
of objects from shadowgraphs with high contrast. Pattern Recognit.
1975;7:157–65.
4. Hounsfield GN: A method and apparatus for examination of a body by
radiation such as x or gamma radiation. UK patent 1283915. 1972.
5. Hounsfield GN: Computerized transverse axial scanning tomography,
part I: description of the system. Br J Radiol. 1973;46:1016–22.
6. Ledley RS, Di Chiro G, Luessenhop AJ, Twigg HL: Computerized
transaxial X-ray tomography of the human body. Science. 1974;186:
207–12.
7. Pan X, Sidky EY, Vannier M: Why do commercial CT scanners still
employ traditional, filtered back-projection for image reconstruction?
Inverse Probl. 2009;25:123009.
8. Ziegler A, K€ohler T, Proksa R: Noise and resolution in images recon-
structed with FBP and OSC algorithms for CT. Med Phys. 2007;34:
585–98.
9. Baek J, Pelc NJ: The noise power spectrum in CT with direct fan beam
reconstruction. Med Phys. 2010;37:2074–81.
10. Gupta AK, Nelson RC, Johnson GA, Paulson EK, Delong DM,
Yoshizumi TT: Optimization of eight-element multi-detector row hel-
ical CT technology for evaluation of the abdomen. Radiology. 2003;
227:739–45.
11. Dictionary.com. Iteration. http://dictionary.reference.com/browse/iteration.
Accessed November 1, 2010.
12. Herman GT, Lent A: Iterative reconstruction algorithms. Comput Biol
Med. 1976;5:319–32.
13. Bittencourt MS, Schmidt B, Seltmann M, Muschiol G, Ropers D,
Daniel WG, Achenbach S: Iterative reconstruction in image space
(IRIS) in cardiac computed tomography: initial experience [online
ahead of print December 1, 2010]. Int J Cardiovasc Imaging. doi:10.
1007s10554-010-9756-3.
14. Silva AC, Lawder HJ, Hara A, Kujak J, Pavlicek W: Innovations in CT
dose reduction strategy: application of the adaptive statistical iterative
reconstruction algorithm. AJR Am J Roentgenol. 2010;194:191–9.
15. Herbert B, Raupach R, Sunnegardh J, Stierstorfer K, Flohr T: Translation
of statistical iterative reconstruction into non-linear image processing.
Presented at the 96th Annual Meeting and Scientific Assembly of the Ra-
diological Society of North America, Chicaco, IL, December 2010.
16. Leipsic J, Labounty TM, Heilbron B, Min JK, Mancini GB, Lin FY,
Taylor C, Dunning A, Earls JP: Adaptive statistical iterative recon-
struction: assessment of image noise and image quality in coronary
CT angiography. AJR Am J Roentgenol. 2010;195:649–54.
17. Leipsic J, Labounty TM, Heilbron B, Min JK, Mancini GB, Lin FY,
Taylor C, Dunning A, Earls JP: Estimated radiation dose reduction us-
ing adaptive statistical iterative reconstruction in coronary CT angiog-
raphy: the ERASIR study. AJR Am J Roentgenol. 2010;195:655–60.
18. Gosling O, Loader R, Venables P, Roobottom C, Rowles N,
Bellenger N, Morgan-Hughes G: A comparison of radiation doses be-
tween state-of-the-art multislice CT coronary angiography with itera-
tive reconstruction, multislice CT coronary angiography with standard
filtered back-projection and invasive diagnostic coronary angiography.
Heart. 2010;96:922–6.
19. Prakash P, Kalra MK, Ackman JB, Digumarthy SR, Hsieh J, Do S,
Shepard JA, Gilman MD: Diffuse lung disease: CT of the chest with
adaptive statistical iterative reconstruction technique. Radiology.
2010;256:261–9.
20. Singh S, Kalra MK, Hsieh J, Licato PE, Do S, Pien HH, Blake MA:
Abdominal CT: comparison of adaptive statistical iterative and filtered
back projection reconstruction techniques. Radiology. 2010;257:373–83.
21. Sagara Y, Hara A, Pavlicek W, Silva AC, Paden RG, Wu Q: Abdom-
inal CT: comparison of low-dose CT with adaptive statistical iterative
reconstruction and routine-dose CT with filtered back projection in 53
patients. AJR Am J Roentgenol. 2010;195:713–9.
22. Prakash P, Kalra MK, Kambadakone AK, Pien H, Hsieh J, Blake MA,
Sahani DV: Reducing abdominal CT radiation dose with adaptive statis-
tical iterative reconstruction technique. Invest Radiol. 2010;45:202–10.
23. Yu Z, Thibault J, Bouman C, Sauer K, Hsieh J: fast model-based x-ray
CT reconstruction using spatially non-homogeneous ICD optimiza-
tion. IEEE Trans Image Process. 2010;99:1.
24. Thibault JB, Sauer KD, Bouman CA, Hsieh J: A three-dimensional
statistical approach to improved image quality for multislice helical
CT. Med Phys. 2007;34:4526–44.