Received: 31 October 2022 Revised: 13 July 2023 Accepted: 13 July 2023

DOI: 10.1002/mp.16679

RESEARCH ARTICLE

Development, validation, and simplification

of a scanner-specific CT simulator

Sjoerd A. M. Tunissen1 Luuk J. Oostveen1 Nikita Moriakov1,2

Jonas Teuwen1,3 Koen Michielsen1 Ewoud J. Smit1 Ioannis Sechopoulos1,4,5

1
Department of Medical Imaging, Abstract
Radboudumc, Nijmegen, The Netherlands
Background: Simulated computed tomography (CT) images allow for knowl-
2
Department of Radiation Oncology, edge of the underlying ground truth and for easy variation of imaging conditions,
Netherlands Cancer Institute, Amsterdam,
The Netherlands
making them ideal for testing and optimization of new applications or algo-
3
rithms. However, simulating all processes that affect CT images can result in
AI for Oncology, Netherlands Cancer
Institute, Amsterdam, The Netherlands simulations that are demanding in terms of processing time and computer mem-
4
Dutch Expert Centre for Screening (LRCB),
ory. Therefore, it is of interest to determine how much the simulation can be
Nijmegen, The Netherlands simplified while still achieving realistic results.
5
Technical Medicine Centre, University of
Purpose: To develop a scanner-specific CT simulation using physics-based
Twente, Enschede, The Netherlands simulations for the position-dependent effects and shift-invariant image corrup-
tion methods for the detector effects. And to investigate the impact on image
Correspondence realism of introducing simplifications in the simulation process that lead to faster
Ioannis Sechopoulos, Department of Medical
Imaging, Radboud University Medical Center,
and less memory-demanding simulations.
Postbus 9101, Geert Grooteplein Zuid 10 Methods: To make the simulator realistic and scanner-specific, the spatial
(route 767), 6525 GA Nijmegen, The resolution and noise characteristics, and the exposure-to-detector output rela-
Netherlands.
Email: [email protected]
tionship of a clinical CT system were determined. The simulator includes a
finite focal spot size, raytracing of the digital phantom, gantry rotation during
Funding information projection acquisition, and finite detector element size. Previously published
FILTER, Grant/Award Number: LSHM19020 spectral models were used to model the spectrum for the given tube voltage.
The integrated energy at each element of the detector was calculated using
the Beer–Lambert law. The resulting angular projections were subsequently cor-
rupted by the detector modulation transfer function (MTF), and by addition of
noise according to the noise power spectrum (NPS) and signal mean-variance
relationship, which were measured for different scanner settings. The simulated
sinograms were reconstructed on the clinical CT system and compared to real
CT images in terms of CT numbers, noise magnitude using the standard devi-
ation, noise frequency content using the NPS, and spatial resolution using the
MTF throughout the field of view (FOV). The CT numbers were validated using
a multi-energy CT phantom, the noise magnitude and frequency were validated
with a water phantom, and the spatial resolution was validated with a tungsten
wire. These metrics were compared at multiple scanner settings, and locations
in the FOV. Once validated, the simulation was simplified by reducing the level
of subsampling of the focal spot area, rotation and of detector pixel size, and
the changes in MTFs were analyzed.
Results: The average relative errors for spatial resolution within and across
image slices, noise magnitude, and noise frequency content within and across

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any
medium, provided the original work is properly cited and is not used for commercial purposes.
© 2023 The Authors. Medical Physics published by Wiley Periodicals LLC on behalf of American Association of Physicists in Medicine.

Med Phys. 2024;51:2081–2095. wileyonlinelibrary.com/journal/mp 2081

2082 SCANNER-SPECIFIC CT SIMULATOR

slices were 3.4%, 3.3%, 4.9%, 3.9%, and 6.2%, respectively. The average abso-
lute difference in CT numbers was 10.2 HU and the maximum was 22.5 HU. The
simulation simplification showed that all subsampling can be avoided, except for
angular, while the error in frequency at 10% MTF would be maximum 16.3%.
Conclusion: The simulation of a scanner-specific CT allows for the generation
of realistic CT images by combining physics-based simulations for the position-
dependent effects and image-corruption methods for the shift-invariant ones.
Together with the available ground truth of the digital phantom, it results in a
useful tool to perform quantitative analysis of reconstruction or post-processing
algorithms.Some simulation simplifications allow for reduced time and computer
power requirements with minimal loss of realism.

KEYWORDS
computer simulations, CT, system characterization

1 INTRODUCTION for specific tasks it could be beneficial to evaluate the

impact on realism of different simplifications that lead to
Currently, computed tomography (CT) is the workhorse substantially shorter computation time.
imaging modality in most radiology departments.1–4 Therefore, we aim to develop a simulation of a
CT is used for screening, diagnosis, and interven- scanner-specific CT capable of generating CT images
tional procedures, such as CT-guided biopsies or with realistic appearance considering the spatial res-
ablations.5–9 Therefore, research on reconstruction and olution and noise characteristics of a clinical CT
post-processing algorithms to increase image quality in system.18,19 With this scanner-specific CT simulation, a
CT, without increasing patient dose, is a growing field broad range of scanner parameters can be simulated,
of interest. Examples of these efforts include develop- such as different tube currents, tube voltages, exposure
ments in deep learning reconstruction of low dose CT,10 times, bowtie filters, and focal spot sizes. In addition,
in denoising of low dose CT using Convolution Neu- once a full simulator is developed and validated, we aim
ral Networks,11,12 and in CT denoising using statistical to evaluate the impact on the realism of the resulting
methods.13,14 images when different simplifications (e.g., less sub-
New reconstruction or post processing algorithms sampling) of the simulator are introduced, with the aim
are typically developed and validated using physical to reduce the time and computer power necessary to
phantoms, which are limited in their capability to rep- simulate an image.
resent real human anatomy. This limits the usefulness
of these phantoms when developing new algorithms.
Patient images can also be used during development 2 METHODS
and validation, but aside from the ethical issues if
new research-specific acquisitions are needed, patient The scanner to be simulated in this work is a 320-
images do not have a quantitative ground truth avail- row CT system (Aquilion ONE PRISM Edition, Canon
able, making it hard to quantify the performance of the Medical Systems, Otawara, Japan) installed at the Dept.
developed algorithms.15 of Medical Imaging of Radboudumc, Nijmegen, The
Therefore, it would be beneficial to have the possibility Netherlands. The scanner-specific CT simulation con-
to test these algorithms using virtual clinical trials.16 In sists of a pipeline (Figure 1) to generate CT images
these, computer simulated images are generated from that are realistic and scanner-specific,based on physics-
digital models of humans, such as the XCAT phantom.17 based simulations for the position-dependent effects
These phantoms have ground truth available and and image corruption methods proposed by Saunders
include considerably realistic anatomy, making them et al. for shift-invariant corruption of ideal images.20
ideal for quantitative evaluation of clinically-relevant In this way the need for proprietary system-specific
conditions. In addition, an infinite number of different information from the vendor is minimized. However,
realizations of the phantoms can be generated and infi- for many of these steps, specific system character-
nite combinations of imaging conditions/parameters can istics need to be known, which, for this work, were
be evaluated, thus the amount of data that can be used obtained via measurements. The details of all steps
for a virtual clinical trial is only limited by computation and these measurements will be discussed in the
time and memory. next sections. A general simulation pipeline, with an
However, simulating a fully detailed CT image that overview of required information for each step, can
incorporates all acquisition process characteristics be found in Appendix A (see online supplemental
accurately is time and memory consuming. Therefore, material).
SCANNER-SPECIFIC CT SIMULATOR 2083

FIGURE 1 Scanner-specific computed tomography (CT) simulation pipeline.

2.1 3D raytracing

To perform the 3D raytracing, the specific geometry

and dimensions of the clinical CT system, including the
detector pixel size, detector distance, focal spot size,
focal spot angle, and focal spot distance, were used.
However, these are vendor-specific and confidential, so
they are not reported here. For the CT system being sim-
ulated, the detector consists of 896 detector channels
and 320 rows. The detector is curved such that all pix-
F I G U R E 2 Schematic drawing of the focal spot and detector
els in a row have the same distance to the source. The elements subsampling.
input to the 3D raytracing is a voxelized phantom, rep-
resenting the object that will be imaged, with the voxel
values indicating an index linking it to the material it con- in the lateral direction from the central ray, in steps of
tains. The 3D raytracing is performed for every material 5 mm out to 160 mm from the CoR, while the tube
present in the phantom, resulting in a separate thickness remained static. The resulting air kerma measurements
map (T) for each material. The raytracing algorithm is a were used to estimate the equivalent thickness of the
GPU-based pixel-driven raytracing based on the work of bowtie filter.
Moriakov et al.21 and Syben et al.22
To account for the finite size of the focal spot, the ray-
tracing is not performed from a single point on the focal 2.2 Primary projection images
spot, but from LxL subsamples of the focal spot, ordered
in a square grid, because the shape of the focal spot To calculate the incident primary photon energy that is
is approximately square. To minimize the discretization absorbed by the detector, first the x-ray spectrum (N)
effect on the detector, the detector elements are sub- leaving the source must be determined. The x-ray spec-
sampled by MxM accordingly and are up-sampled later trum is modeled by measuring the incident air kerma
in the simulation process.20 To incorporate the effect of as close as possible to the detector (8 cm away) under
the finite exposure time during the angular motion of four different attenuation conditions: no added attenu-
the CT gantry that causes spatial resolution loss, the ating materials, 6 mm aluminum, and 1 mm and 2 mm
angular projections are also subsampled by a factor K. copper. These air kerma measurements are used to fit
The focal spot and detector subsampling is depicted the spectrum model.24 The same measurements and
in Figure 2 and the angular subsampling is depicted in attenuations are used to obtain the conversion from
Figure 3. absorbed primary photon energy to digital units (DU).
This results in thickness maps containing the inter- This conversion was determined for each bowtie filter
section length for each material from each focal spot separately.
subsample fa,b to the center of each subsampled As depicted in Figure 1, the x-ray spectrum is used
detector element (xi , yj ) at each subsampled angular together with the thickness map of each material to
projection 𝜃k . determine the primary photon energy absorbed by the
To include the effect of the bowtie filters in the sim- detector, that is, the simulated sinogram I(fa,b , xi , yj , 𝜃k )
ulation, the shape of the bowtie filters was determined according to the Beer–Lambert law25 :
( )
using the method described by McKenney et al.23 For I fa,b , xi , yj , 𝜃k
this, the air kerma was measured with a dosimeter ( )
∑ ∑
(10 × 6-0.6CT, Radcal, Monrovia, California, USA) at the = e∗Ne ∗QEe ∗ exp − 𝜇m,e Tm,fa,b ,xi ,yj ,𝜃k (1)
center of rotation (CoR) of the CT gantry and outwards e m
2084 SCANNER-SPECIFIC CT SIMULATOR

FIGURE 3 Schematic drawing of the angular subsampling.

where Ne is the number of photons with energy e emit- The other causes of spatial resolution loss, namely,
ted from the source, QEe is the quantum efficiency of the focal spot size effect and the blur caused by expo-
the detector for each energy e, 𝜇m,e is the attenua- sure time per angular projection, are already included in
tion coefficient of material m at energy e (determined the image, as described above, by raytracing the focal
using the xraydb package in python based on the work spot and angular projections, including subsampling. To
of Elam et al.,26 and the work of Boone et al.27 ), and maximize the realism of the image simulation, these
Tm,fa,b ,xi ,yj ,𝜃k is the thickness map of each material m at effects must be included in the raytracing step, since
each subsampled detector element (xi , yj ) for each focal they are position-dependent in the field of view (FOV),
spot subsample fa,b and at each subsampled projection and therefore this information cannot be added to the
angle 𝜃k . sinogram directly.
After the incorporation of the detector MTF, the sim-
ulated sinogram is binned to its real dimension, using
2.3 Spatial resolution loss Equation (2):

The spatial resolution characteristics of the detector are 1 ∑M ∑ M ∑ K ∑ L

incorporated by applying the modulation transfer func- I (x, y, 𝜃) =
M 2 ∗ K ∗ L2 i = 1 j = 1 k = 1 a = 1
tion (MTF) of the detector to the simulated sinogram
resulting from Equation (1). This is done by multipli-
cation, in frequency domain, of the MTF with the 2D ∑
L
( )
I fa,b , xi , yj , 𝜃k (2)
fast Fourier transform (FFT) of the simulated sinogram b=1
(at each subsampled projection angle) and then taking
the inverse FFT. Please note that the MTF needs to be where I(x, y, 𝜃) is the sinogram after all subsamples are
divided by the sinc function of the final detector spacing, averaged at detector pixel x, y and angular projection 𝜃,
since the sampling of the detector causes the MTF to M is the number of detector pixel subsamples in each
be multiplied by the sinc function. direction, K is the number of angular subsamples, L is
The detector MTF was measured in the detector row the number of focal spot subsamples in each direction,
direction, since this direction does not suffer from reso- I(fa,b , xi , yj , 𝜃k ) is the sinogram with all subsamples, fa,b
lution loss due to rotation, and was used in all directions, is the focal spot subsample, xi , yj is the detector pixel
assuming it is rotationally invariant. The slanted edge subsample, and 𝜃k is the angular projection subsample.
method28 was used with a tungsten edge (TX5, IBA
Dosimetry, Schwarzenbruck, Germany). The tungsten
edge was placed as close as possible to the detector 2.4 Noise addition
(8 cm away) to minimize the focal spot size effect. A
Lorentzian based fit29 is used to fit the MTF. The mea- To add the correct noise to the sinogram two character-
sured edge and fitted MTF are shown in Appendix B. istics of the noise need to be known, the mean-variance
The fit could potentially result in values close to the zero relationship of the noise signal and the noise power
frequency to be larger than one. These are forced to one spectrum (NPS). Both are determined using the same
when applying the MTF. scans of two water phantoms of 240 mm and 320 mm
SCANNER-SPECIFIC CT SIMULATOR 2085

in diameter, representing the attenuation of brain and image with the correct NPS, 𝜎( −1 {N(u, v)}) is the stan-
abdomen, respectively. dard deviation and 𝜇( −1 {N(u, v)}) the mean of the
The mean and variance of the signal were deter- colored noise in image domain after inverse FFT, MV is
mined in a 20 × 30-pixel region of interest (ROI) at the mean-variance relationship, I is the primary projec-
approximately the center of each sinogram projection, tion image after the MTF is applied and binned and Inoise
and their averages over all projections were used as is the projection image after the noise is added to it.
the final mean and variance. To obtain the NPSs, first
a correction for image lines due to detector tiling was
performed by averaging all projections and subtracting 2.5 Hounsfield unit calibration
the result from each individual projection. Second, the
2D FFT of a 64 × 64-pixel ROI in the center of the sino- As is standard in CT imaging, a linear calibration was
gram was calculated for each projection and the square obtained to apply to all reconstructed images to cor-
of the absolute value of these FFTs was taken. The rect the resulting CT numbers for different materials
results were averaged for all projections, resulting in 2D and densities. The applied linear correction was deter-
NPSs. Despite the anisotropic pixel size there was no mined by digitally simulating and physically measuring a
significant difference between the NPS in the horizontal cylindrical water phantom with 5 different inserts: Teflon,
and vertical directions, so they were radially averaged Delrin, acrylic, polypropylene, and air (quality control
to obtain a 1D NPS. The mean-variance relationship is phantom provided by Canon Medical Systems) and fit-
dependent on the tube voltage, and bowtie filter. The ting the CT numbers of the simulation to the physically
shape of the NPS is dependent on the tube current, measured CT numbers. The diameter of this phan-
tube voltage, exposure time, and bowtie filter. Hence, tom is 190 mm, and the inserts have a diameter of
both were measured at nine different tube current lev- 20 mm. A linear correction was obtained from the mean
els between 10 and 400 mA, four different tube voltage HU of the simulated phantom inserts and water back-
levels, 80, 100, 120, and 135 kV, two different exposure ground and their corresponding theoretical values, using
times 0.275 and 0.5 s, and for two different bowtie filters. Equation (6).
The mean-variance relationship (MV) is defined as a
linear function with a positive offset (Equation 3). This min
2
(HUmeasurement − a ∗ (HUtheoretical + bwater ))
offset is the electronic noise. a,bwater ∈ℝ
(6)
MV = a ∗ m + belectronic noise (3) The value of bwater (offset of water) was fit such
that the simulated water value corresponds to the the-
where a is the slope of MV, m is the mean in a 3 × 3 pixel oretical one, that is, equal to zero. Afterwards a (slope)
region (Saunders et al.20 ), and belectronic noise is the offset was fit such that the HUtheoretical , after correction, had
due to the electronic noise. the smallest possible error against the corresponding
To get the desired noise, white noise is generated with HUtheoretical .This was done separately for each tube volt-
similar spatial dimensions as the sinogram projections, age level available in the system, and the corresponding
as described by Saunders et al.20 The resulting noise is calibration was then applied to all subsequent simulated
multiplied in frequency domain with the square root of images.
the NPS (Equation 4) and is scaled by the MV in spatial
domain to obtain the desired noise, which is then added
to the sinogram I(x, y, 𝜃) (Equation 5). 2.6 Validation of simulation
√ To assess the accuracy of the simulator, multiple vali-
N (u, v) = NPS ∗ {n (𝜇 = 0, 𝜎 = 1)} (4) dations were performed to validate the CT numbers of
different materials, the spatial resolution, and the noise
√ characteristics of the simulated images against images
Inoise (x, y, 𝜃) = I (x, y, 𝜃) + MV (I (x, y, 𝜃))
acquired with the clinical CT system. All validations were
(
1 ( performed after reconstruction of the sinogram projec-
∗ ( ) ∗  −1 {N (u, v)} tions on the clinical CT system using the clinically avail-
𝜎  −1 {N (u, v)}
able filtered back projection (FBP), which is based on
) the Feldkamp Davis Kress (FDK) algorithm.30 For the CT
( −1 ))
−𝜇  {N (u, v)} (5) number and noise characteristics validation, the number
of angular subsamples K was set to 2, the number of
focal spot subsamples L was set to 1, and the number
where  is the FFT operator, n is a realization of white of detector subsamples M was set to 2. For the resolu-
Gaussian noise with mean 𝜇 and standard deviation tion loss validation, the number of angular subsamples
𝜎,  −1 is the inverse FFT, N(u, v) is the colored noise K was set to 3, the number of focal spot subsamples L
2086 SCANNER-SPECIFIC CT SIMULATOR

was set to 3, and the number of detector subsamples M (SSP) was determined in the same way as the MTF. The
was set to 4. These subsampling factors were obtained simulated voxel size of the phantom was 0.005 mm ×
experimentally, the details can be found in Appendix C. 0.035 mm × 0.005 mm.The voxel size in the lateral direc-
The CT numbers were validated using a physical oval tion was substantially larger and set to this exact value
phantom (with 40 and 30 cm radii for the horizontal because with a shift of one pixel per lateral (the direc-
and vertical directions, respectively) with 15 cylindri- tion with pixel size 0.035 mm) step this results in the
cal inserts, each of different material and of diameter simulated wire being placed at an angle of 8 degrees.
28.5 mm31 (MECT phantom, Sun Nuclear, Middleton, Please note that the simulations were noiseless, since
WI, USA). The exact dimensions and material compo- noise does not influence the resolution loss.
sition of the MECT phantom were known, so we could The 50 μm diameter of the tungsten wire is rela-
not only image but also simulate the phantom and its tively small compared to the detector pixel size, even
image acquisition, with a tube current of 400 mA and when subsampled. To overcome this problem the detec-
three different tube voltage levels (100,120,and 135 kV). tor subsampling M was set to 24, just for the raytracing.
The simulated voxel size of the phantom was 3.3 mm × After the raytracing, the detector was rebinned to its
0.25 mm × 0.25 mm. The voxel size in the longitudinal original subsampling of M = 4.
direction was substantially larger since the phantom is For validating the noise magnitude and frequency
constant in this direction. The measured and simulated content, a water phantom with a radius of 320 mm was
sinograms were both reconstructed on the clinical CT again both digitally simulated and physically measured,
system using FBP and a FOV of 320 mm × 320 mm and the results were compared. The noise magnitude
and 160 mm in the longitudinal direction. The recon- and frequency content were validated at two different
structed volume consisted of 320 slices of 512 × 512 tube current levels (140 and 400 mA), and three different
pixels.The hounsfield units (HU) within these inserts and tube voltage levels (100, 120, and 135 kV). The simu-
in the water-equivalent background was measured by lated voxel size of the phantom was 1.0 mm × 0.25 mm
averaging a squared 10 × 10 pixel ROI across 80 slices. × 0.25 mm. The voxel size in the longitudinal direction
The resolution loss of the simulator was validated was substantially larger since the phantom is constant in
by imaging a 50 μm diameter tungsten wire,32,33 both this direction. A volume of interest (VOI) of 64 × 64 × 64
digitally and physically. This tungsten wire creates a voxels was placed in the center of the water phantom
Dirac delta function or unit impulse,34 and the point- images. The standard deviation of this VOI was used to
spread function (PSF) is obtained by taking the Radon validate the magnitude of the noise.To validate the noise
transform35 of this signal in one direction. The MTF frequency content, a 100 mm FOV was reconstructed in
is then determined by calculating the FFT of the PSF. the center and at the periphery, approximately 120 mm
The spatial resolution was validated at 7, 14, and from the center, of the water phantom. This smaller FOV
21 cm from the isocenter, for both the digitally sim- was reconstructed, to have a smaller pixel size, making
ulated wire and the real physical measured wire to it possible to validate higher frequencies. The 2D NPS
verify the validity of the simulation of the shift-variant and 2D unstructured NPS of both these FOVs, were cal-
rotational blur and focal spot size effects. Each sim- culated in 256 × 256-pixel ROIs from across 80 slices,
ulated and measured wire was reconstructed with a by determining the square of the 2D Fourier transform.
small FOV of 19.5 mm × 19.5 mm of 512 × 512 pix- In the case of the unstructured NPS the average of the
els, so the PSF had enough samples. The resolution 80 slices was subtracted before calculating the Fourier
loss was checked for both focal spot sizes present in transform. Both the 2D NPS and 2D unstructured NPS
the clinical system, which will be referred to as large were normalized to have an area of one, obtaining the
and small focal spots from here on, and for both the normalized NPS (nNPS). Both the 2D nNPS and 2D
radial and tangential direction for all positions. The sim- unstructured nNPS were calculated to show that the
ulated voxel size of the phantom was 0.1432 mm × simulation does not introduce any structured noise. A
0.005 mm × 0.005 mm. The voxel size in the longitu- comparison of the nNPS at these two positions was
dinal direction was substantially larger and set to this performed to validate the changes in the noise charac-
exact value because with a shift of one pixel per longitu- teristics throughout the imaging field. These 2D nNPSs
dinal (the direction with pixel size 0.1432 mm) step this at the center were also radially averaged, and again nor-
results in the simulated wire being placed at an angle malized to have an area of one, to obtain a 1D nNPS.
of 3 degrees. Please note that the simulations were To validate the frequency content across slices, the 1D
noiseless, since noise does not influence the resolution nNPS was calculated across 280 slices for all pixels in
loss. a 128 × 128 ROI at the center, and the results were
The resolution loss in longitudinal direction (across averaged.
slices) has also been validated by imaging this 50 μm Scatter was not included in our simulator, since the
diameter of the tungsten wire both digitally and phys- system performs scatter correction during the recon-
ically. The wire was placed such that the angle with struction process, and therefore, the benefit of adding
the slices was 8 degrees. The slice sensitivity profile simulated scatter would be minimal. To validate the
SCANNER-SPECIFIC CT SIMULATOR 2087

performance of the scatter correction, the 320 mm water

phantom (also used for the validation of the noise) was
imaged with the standard volume scan collimation of
160 mm (equal to all measurements in this work) and
with a 20 mm collimation, which is assumed to have
a negligible amount of scatter. The line profile of the
reconstructed water phantom images was compared for
both collimations. Line profiles were obtained from these
images by averaging 60 individual line profiles across
38 slices both horizontal and vertical directions for both
water phantom scans and the corresponding simulation.

2.7 Simulation simplifications F I G U R E 4 MECT phantom, measured with 135 kV, used for
computed tomography (CT) number validation, with a window level
The three steps of the simulator incorporating sub- (WL) of 200 HU and a window width (WW) of 1000 HU.
sampling, namely, the number of angular projection
subsamples K, the number of detector subsamples M ×
M and the number of focal spot subsamples L × L, were
simplified to reduce the time and computer power nec-
essary. The angular subsamples K were set to 1, 2, and
3. The detector subsamples M × M were set to 1 × 1,
2 × 2, 3 × 3 and 4 × 4. The number of focal spot sub-
samples L × L was set to 1 × 1, 2 × 2, and 3 × 3. Please
note that while one of these three was reduced the
other two were kept at their original value. Previous CT
simulators18,36,37 also used or optimized their subsam-
pling, however with this analysis the impact of each
individual simplification is shown. F I G U R E 5 Computed tomography (CT) numbers of the real and
simulated MECT phantom images at 135 kV. Note that the B stands
To validate the impact of these simplifications on the
for blood ρ = 1.03 g/cm.3
realism of the simulation, the MTFs of the images result-
ing from the digitally simulated simplified sinograms
were determined and compared to the MTFs of the HU. As can be seen from the values in Table 1, the simu-
physically measured sinograms. In both cases these lations result in a small negative bias in the CT numbers
MTFs were again determined from a tungsten wire at for all tube voltage levels.
7, 14, and 21 cm from the isocenter. Figure 5 shows the measured and simulated CT num-
In addition, the possibility of compensating for sim- bers of the 135 kV case. The remaining errors in the
plifying the focal spot as being a point source by using CT numbers do not seem to have a correlation, indi-
the system MTF, that is, the MTF measured with the cating the simulation does not introduce any non-linear
edge located at the CoR, instead of the detector MTF, offset to the CT numbers. Note that, as explained before,
was also tested. These MTFs are shown in Appendix the linear calibration is determined using a different
B. Finally, the possibility of simplifying the incorpo- phantom and is applied to all subsequent images.
ration of the rotational blurring, due to the angular The MTFs in the radial and tangential directions can
motion of the source and detector, was also investigated be seen in Figure 6. It can be observed that the latter
by averaging each angular projection with the subse- starts dropping when moving out of the CoR in both the
quent one, instead of performing the angular projection measured and simulated cases. This is due to the rota-
subsampling. tion of the system introducing more blur further away
from the CoR. The frequencies at 10% MTF and their
relative error for the various FoV positions, directions,
3 RESULTS and focal spot sizes are listed in Table 2. The maximum
error is 11.1% and the mean absolute error is 3.4%,
In Figure 4, one of the measurements of the MECT showing that the spatial resolution characteristics in the
phantom used for the CT number validation is shown. simulated CT images are close to those of the clinical
Table 1 shows the measured and simulated CT numbers system.
for the different materials in the MECT phantom, with The SSP of both the large and small focal spot
the numbers corresponding to the regions in Figure 4. can be seen in Figure 7. It can be observed that the
Table 1 also shows the maximum (bold and underlined), resolution loss is higher for the large focal spot, as
mean absolute, and mean error of the CT numbers in expected.Table 3 shows the frequencies at 10% SSP for
2088 SCANNER-SPECIFIC CT SIMULATOR

TA B L E 1 Computed tomography (CT) numbers corresponding to the regions indicated in the MECT phantom of Figure 4. The material with
the maximum error in CT number is bold and underlined for each tube voltage. Note that the B in material 11 and 12 stands for blood with
ρ = 1.03 g/cm3 . The error is defined as simulation − measurement.

100 kV 120 kV 135 kV

Measurement Simulation Measurement Simulation Measurement Simulation
Material [HU] [HU] [HU] [HU] [HU] [HU]

(1) Solid water 8.1 −0.6 3.9 −5.3 4.6 −4.0

(2) Adipose −61.4 −74.7 −52.5 −69.7 −48.8 −66.2
(3) Brain 43.4 28.0 42.7 26.6 42.3 26.7
(4) 2 mg/mL I 63.6 54.1 51.6 38.6 48.3 35.4
(5) 5 mg/mL I 143.9 132.3 114.6 103.7 103.4 90.8
(6) 10 mg/mL I 267.4 263.5 206.2 203.8 178.0 177.5
(7) 15 mg/mL I 384.1 385.4 296.5 303.3 258.0 264.1
(8) Blood ρ = 1.03 54.8 41.8 45.5 38.8 44.8 41.3
(9) Blood ρ = 1.07 78.1 65.6 70.9 63.7 70.6 68.2
(10) Blood ρ = 1.1 106.0 97.4 104.7 97.3 106.3 100.3
(11) 2 mg/mL I + B 107.6 95.6 92.4 82.1 86.5 77.3
(12) 4 mg/mL I + B 149.6 140.2 127.0 119.6 117.8 111.2
(13) 50 mg/mL Ca 189.8 169.9 176.3 162.4 171.2 157.7
(14) 100 mg/mL Ca 335.0 313.7 302.4 281.7 291.6 269.1
(15) 300 mg/mL Ca 882.9 886.3 776.3 779.6 728.9 732.8
Mean absolute error [HU] 10.9 10.2 9.4
Mean error [HU] −10.3 −8.8 −8.1

FIGURE 6 Modulation transfer function of measured and simulated wires in radial (left) and tangential (right) direction for the large focal
spot.

measurement and simulation and their relative error for The ROIs used to validate the frequency content
both focal spots.The mean absolute error is 3.3%,show- within a slice are also indicated in Figure 8, by the red
ing that the spatial resolution across slices in the simu- and yellow squares. Please note that the ROIs of the
lated CT images is close to those of the clinical system. measurements have a slight offset in vertical direction,
Figure 8 shows images of the water phantom used this offset compensates for misalignment between the
to validate the noise characteristics of the simulated measured and simulated water phantom, to ensure that
images. The red square indicates the ROI used for the the same location of the phantom was analyzed. The
validation of the noise magnitude. Table 4 shows the 2D nNPS and 2D unstructured nNPS of the noise at the
results of the noise magnitude in terms of standard devi- center (red square in Figure 8) within a slice of the mea-
ation. The maximum and mean absolute errors were sured and simulated phantom are shown in Figures 9
8.5% and 4.9%. and 11, respectively. It can be seen that the 2D nNPS
SCANNER-SPECIFIC CT SIMULATOR 2089

TA B L E 2 Frequency at 10% MTF of the measured and simulated MTF, and the relative error of these frequencies in the different directions,
positions, and focal spot sizes evaluated. Relative error = (simulation − measurement)/measurement * 100%.

Measured Simulated
Focal spot size Direction Position (cm) frequency (mm−1 ) frequency (mm−1 ) Relative error (%)

Large Tangential 7 0.62 0.62 0.3

14 0.57 0.57 1.2
21 0.49 0.51 3.6
Radial 7 0.62 0.65 3.8
14 0.56 0.63 11.1
21 0.37 0.38 2.6
Small Tangential 7 0.72 0.70 −3.4
14 0.64 0.63 −1.5
21 0.55 0.55 0.8
Radial 7 0.77 0.74 −3.3
14 0.70 0.73 3.6
21 0.41 0.39 −5.7

and 2D unstructured nNPS are isotropic in the center

for both measurement and simulation. The difference
images only show a small overestimation at lower fre-
quencies (white dominant ring) and underestimation at
slightly higher frequencies (black dominant ring). The
2D nNPS and 2D unstructured nNPS of the noise at
the periphery (yellow square in Figure 8) of the mea-
sured and simulated phantom are shown in Figures 10
and 12, respectively. These results show that the nNPS
is anisotropic at the periphery for both measurement
F I G U R E 7 Slice sensitivity profile of measured and simulated and simulation, and that the degree of anisotropy in
wires for both focal spots present in the system. the simulated image is similar to that in the real one.
The difference images only show a small overestima-
TA B L E 3 Frequency at 10% SSP of the measured and tion in vertical direction (white dominant regions above
simulated SSP, and the relative error of these frequencies for both and below the center) and underestimation in horizon-
focal spot sizes evaluated. Relative error = (simulation − tal direction (black dominant regions left and right from
measurement)/measurement * 100%.
the center). The only difference between the 2D normal-
Measured Simulated ized nNPSs and 2D unstructured normalized nNPSs is
Focal spot frequency frequency Relative a small low frequency peak in the horizontal direction of
size (mm−1 ) (mm−1 ) error (%)
the 2D nNPS of the periphery, which is not present in the
Large 1.14 1.12 −1.7 2D unstructured nNPS of the periphery. This minor peak
Small 1.48 1.55 4.8 is introduced by a small cupping artifact in the periphery
of the measurement, see Figure 15.
TA B L E 4 Noise magnitude of measured and simulated water Figure 13 shows the radially averaged nNPS and
phantom of Figure 8. Relative error = (simulation − nNPS across slices in the center of both the mea-
measurement)/measurement * 100%. surement and simulation for the 135 kV and 140 mA
Tube σ measure- case. The mean absolute errors of radially averaged
Tube current ment σ simulation Relative nNPSs are summarized in Table 5, which shows that
voltage (kV) (mA) (HU) (HU) error (%) the maximum mean absolute error is 8.4% and the
100 140 128.2 140.2 8.5 average mean absolute error is 3.9%. The mean abso-
lute errors of the nNPSs in the slice direction are
400 66.8 68.8 2.9
summarized in Table 5, and is on average 6.2% and
120 140 83.8 89.0 5.8
maximum 8.8%. The plots of the other radially aver-
400 47.0 48.1 2.2 aged nNPSs and nNPSs across slices are depicted in
135 140 68.4 72.8 6.1 Appendix D.
400 39.0 40.4 3.6 Figure 14 indicates the region from where the
line profiles of the water phantom for the scatter
2090 SCANNER-SPECIFIC CT SIMULATOR

F I G U R E 8 Water phantom used for nNPS validation (140 mA, 135 kV) with a WL of 0 HU and a WW of 400 HU. The squares indicate the
ROIs used to determine the nNPSs.

F I G U R E 9 2D nNPS in the center region of measured (left) and simulated (middle) water phantom, and the difference between both
nNPSs (right). Difference = simulation − measurement.

F I G U R E 1 0 2D nNPS in the periphery region of measured (left) and simulated (middle) water phantom, and the difference between both
nNPSs (right). Difference = simulation − measurement.

F I G U R E 1 1 2D unstructured nNPS in the center region of measured (left) and simulated (middle) water phantom, and the difference
between both nNPSs (right). Difference = simulation − measurement.
SCANNER-SPECIFIC CT SIMULATOR 2091

F I G U R E 1 2 2D unstructured nNPS in the periphery region of measured (left) and simulated (middle) water phantom, and the difference
between both nNPSs (right). Difference = simulation − measurement.

FIGURE 13 Radially averaged center nNPS (left) and nNPS across slices (right), for 135 kV and 140 mA.

TA B L E 5 Difference in measured and simulated nNPS with a slice and across slices.

Tube voltage 100 kV 120 kV 135 kV

Tube current 140 mA 400 mA 140 mA 400 mA 140 mA 400 mA

% Mean absolute 2.2 2.2 2.4 4.9 3.1 8.4

difference within slice
% Mean absolute 5.6 5.7 5.2 7.3 4.8 8.8
difference across slices

F I G U R E 1 4 Averaged slices of simulated water phantom with

indicated ROIs used for obtaining the line profiles. F I G U R E 1 5 Line profiles of the measurement with 160 mm
collimation, the measurement with 20 mm collimation, and the
simulation.

correction validation are obtained, with the resulting

average line profiles shown in Figure 15. The line pro- maximum difference of 15 HU at the edge of the water
files show the effect of the scatter correction, resulting phantom.
in a good match in HU values at the center of the phan- Tables 6 and 7 show the absolute relative errors
tom, but with a larger remaining error at the sides. As a between the simplified and the fully subsampled simu-
result, it can be seen that the line profile of our simula- lation of the frequency at 10% MTF for the simulated
tion is also approximately flat, as expected, and has a tungsten wire at 7, 14, and 21 cm from the CoR. The
2092 SCANNER-SPECIFIC CT SIMULATOR

TA B L E 6 Frequency at 10% MTF for all simulation simplifications and absolute relative error of this frequency compared to the frequency
of the full simulation at 10% MTF, for tangential direction.

Absolute relative error at 10%

MTF
Time/memory
Subsampling Subsampling reduction factor
Focal spot size simplification factor 7 cm 14 cm 21 cm

Large Focal spot 1 20.9 18.0 12.1 9

2 7.7 8.5 4.1 2.25
System MTF 1 12.5 11.9 7.6 9
Angular 1 2.5 6.0 16.5 3
2 0.4 0.7 1.8 1.5
Projection averaging 1 15.2 14.5 15.3 3
Detector 1 0.4 0.0 0.2 16
2 0.1 0.0 0.0 4
3 0.1 0.1 0.1 1.78
Small Focal spot 1 6.9 6.6 3.8 9
2 3.0 3.4 2.2 2.25
System MTF 1 0.6 1.1 0.4 9
Angular 1 2.5 10.3 20.1 3
2 0.4 1.5 2.3 1.5
Projection averaging 1 16.9 12.6 9.7 3
Detector 1 0.2 0.0 0.4 16
2 0.0 0.0 0.1 4
3 0.0 0.1 0.1 1.78

TA B L E 7 Frequency at 10% MTF for all simulation simplifications and absolute relative error of this frequency compared to the frequency
of the full simulation at 10% MTF, for radial direction.

Absolute relative error at 10% MTF Time/memory

Focal spot Subsampling Subsampling reduction
size simplification factor 7 cm 14 cm 21 cm
factor
Large Focal spot 1 21.1 25.2 2.8 9
2 11.2 10.4 1.7 2.25
System MTF 1 12.7 16.3 3.4 9
Angular 1 0.1 0.1 0.1 3
2 0.1 0.0 0.0 1.5
Projection averaging 1 0.6 1.0 0.4 3
Detector 1 −0.6 2.1 0.1 16
2 0.2 0.3 0.1 4
3 0.3 0.1 0.1 1.78
Small Focal spot 1 5.5 7.7 0.8 9
2 3.7 3.8 0.5 2.25
System MTF 1 1.8 0.0 1.3 9
Angular 1 0.0 0.1 0.0 3
2 0.1 0.0 0.0 1.5
Projection averaging 1 1.1 1.3 0.1 3
Detector 1 1.2 1.3 0.2 16
2 0.1 0.2 0.0 4
3 0.0 0.1 0.0 1.78

absolute relative errors for all simplifications are shown number of calculations that need to be performed scales
for the tangential (Table 6) and radial (Table 7) direc- linearly with the amount of memory in our application).
tion. In the last column of both Tables 6 and 7, the time The differences in the resulting MTFs are, for most
and memory reduction factor are listed (the reduction cases, subtle, except for the case of not subsam-
factor is the same for time and memory, because the pling the focal spot or angular projections. The detector
SCANNER-SPECIFIC CT SIMULATOR 2093

GPU. All image corruptions were performed on the CPU

(due to the sinogram size). Generating all projections of
896 × 320 pixels for the MECT phantom, with a voxel
array size of 50 × 1600 × 1600 voxels of the same
voxel size as those used for validation and consisting of
15 different materials (used for CT number validation),
took ∼40 h on this workstation. Please note that the time
reduction factor reported with the simplification results is
F I G U R E 1 6 Noiseless simulations of a lesion at 14 cm from the theoretical as some minor operations are independent
CoR. (Left) Full simulation (WW: 100, WL: 85). (Middle) Simplified
of sinogram size and the calculations of an entire set
simulation, 1 source sample using the MTF measured in the CoR, 2
angular subsamples and 1 detector sample, time and memory of projections are done in batches due to memory con-
consumption potentially reduced by a factor 216 (WW: 100, WL: 85). straints. As an example, by reducing the sub-sampling of
(Right) Difference between the two simulations (WW: 40, WL: 0). the focal spot to 1, angular projections to 2, and detector
pixels to 1, the simulation time for this same simulation
results in ∼35 min.
subsampling seems to have very little effect in both
directions.
Tables 6 and 7 also show the absolute relative errors 4 DISCUSSION
in the frequency at 10% MTF between using the sys-
tem MTF (as described in Section 2.7) and the fully In this work a scanner-specific CT simulation was
subsampled focal spot simulation for the simulated tung- developed and validated, combining physics-based
sten wire at 7, 14, and 21 cm from the CoR. The results simulations for the position-dependent effects and the
clearly show that using the system MTF improves the shift-invariant image corruption methods described by
results when assuming the focal spot is a point source, Saunders et al.20 for the detector effects. Therefore, min-
since it reduces the error compared to the full simulation imizing the system information needed from the vendor
by approximately 30%−40%. The results also show that and making it possible to perform scanner-specific
averaging each angular projection with the subsequent CT simulations with only system-specific geometry
angular projection gives worse results than when simpli- information. In addition, the impact of simplifying the
fying to only using a single angular projection. Therefore, simulation process, both in terms of resulting realism
to have an error of 17% or less in the frequency at 10% and computer power requirements, was evaluated. To
MTF compared to the full simulation, all subsamples include all effects of the real CT system resulting in spa-
can be reduced to 1, except for the angular projections, tial resolution loss, the shift-variant impact of the finite
which should still be 2. In this case the maximum error focal spot size and of gantry rotation are modeled. The
is 16.3% and the time and memory consumption could simulator is shown to generate images that match the
be reduced by a factor of 216. characteristics of the real images to within an average
The influence of these simplifications (resulting in of 3.4% and 3.3% in terms of spatial resolution within
16.3% error) can be observed in Figure 16. Here a and across image slices, respectively. In terms of noise
small lesion is imaged with full subsampling and sim- characteristics, these matched in terms of noise magni-
plified subsampling (1 source sample using the system tude (standard deviation) and noise frequency content
MTF, 2 angular subsamples, and 1 detector sample) at (nNPS) within and across slices to within 4.9%, 3.9%,
14 cm from the CoR, to show its influence. The differ- and 6.2%, respectively. The effect of simulation simpli-
ence image shows a minor ring but no other structural fication was assessed, and the results showed that for
differences, indicating minimal difference between the general applications most simplifications, except for the
two simulations after reconstruction. The same can be angular simplification, do not cause a major decrease in
observed from the line profiles of these two recon- realism of the simulated image (maximum error in fre-
structed simulations plotted in Figure 17. However, for quency at 10% MTF of 16.3%). Therefore, the time and
each application, the desired accuracy could differ, so computer power necessary could be reduced for many
the user should decide what is an acceptable error applications in which this level of realism is sufficient.
margin for their application. This would aid studies that aim to use large virtual clin-
ical trials, since it will become feasible to generate very
extensive datasets within a reasonable time frame. Also,
3.1 Computation time studies about processing or reconstruction algorithms
could benefit as it becomes more feasible to cover large
All simulations were performed on a Linux system with multi-dimensional parameter spaces for (first stage)
128 GB RAM, AMD Ryzen Threadripper 1950 × 16-core testing, after which one could choose for more realistic
CPU, and a 48 GB Nvidia RX A600 GPU. The ray trac- simulations for refinement of solutions, if needed. How-
ing and sinogram calculations were performed on the ever, the purpose of the study must be considered to
2094 SCANNER-SPECIFIC CT SIMULATOR

F I G U R E 1 7 (Left) Line profile of the full and simplified simulation of the lesion in Figure 16. (Right) Difference between simplified
simulation and full simulation (simplified simulation − full simulation).

make an informed decision on the level of realism that is 5 CONCLUSION

desired.
During the development some assumptions had to be A scanner-specific CT simulation was developed, imple-
made. One of them being the assumption of stationary mented, and validated. The validation of the performed
behavior of the detector MTF and NPS across the entire simulations showed that it can generate images compa-
detector. The NPSs were only measured at the detector rable to those obtained using a real clinical CT system.
center, since the curvature of the detector is such that An analysis on simulation simplification also showed
the normal direction of each element is pointing to the that for general applications, time and computer power
source. Only in the direction of the rows, where there is can be spared without substantial loss of realism. The
no curvature, the normal is not pointing directly to the simulator can generate realistic scanner-specific CT
source, however the maximum angle is 10◦ and there- images, which will aid the development of new recon-
fore assumed to have a negligible effect. Also, the MTF struction and post-processing algorithms by opening the
was only measured in the detector row direction (since possibility for virtual clinical trials.
in the other direction the MTF is affected by the gantry
rotation), then assuming rotational symmetry between AC K N OW L E D G M E N T S
all directions. In addition, the resolution characteristics We wish to acknowledge Health Holland for funding
are dominated by the focal spot size and finite expo- project FILTER (project LSHM19020) and Canon Med-
sure time during the angular motion of the CT gantry, ical Systems for their assistance in developing this
justifying the assumption of stationary and symmetric simulator.
MTF across the detector. Also, the residual scatter after
correction was neglected. As shown, this simplification C O N F L I C T O F I N T E R E S T S TAT E M E N T
results in a CT number difference of approximately 15 The authors have the following conflicts of interest
HU at the edge of a 320 mm water phantom with the to disclose: Jonas Teuwen—shareholder Ellogon.AI;
widest x-ray beam collimations. The QE of the detec- Ewoud J. Smit—speaker bureau of Canon Medi-
tor is based on the theoretical energy absorption of the cal Systems and patent inventor 4D Similarity Fil-
nominal detector active layer thickness. However, even ter; Ioannis Sechopoulos—research agreements with
with these assumptions, the validation results point to Siemens Healthcare, Canon Medical Systems, Screen-
the appropriateness of the simulations in terms of spa- Point Medical, Sectra Benelux, Volpara Healthcare,
tial resolution, noise magnitude and frequency content, Lunit and iCAD, and speaker agreement with Siemens
and, especially, their shift variance compared to a real Healthcare.
clinical CT system.
The developed scanner-specific CT simulation may REFERENCES
function as a tool to facilitate virtual clinical trials to 1. Hess EP, Haas LR, Shah ND, Stroebel RJ, Denham CR, Swensen
test new reconstruction algorithms and post-processing SJ. Trends in computed tomography utilization rates. J Patient
algorithms. The simulator could additionally be used to Saf. 2014;10(1):52-58. doi:10.2307/26633038
2. Broder J, Fordham LA, Warshauer DM. Increasing utilization
also test new components like bowtie filters or flat filters,
of computed tomography in the pediatric emergency depart-
different detectors, or a different focal spot size, as well ment, 2000–2006. Emerg Radiol. 2007;14(4):227-232. doi:10.
as new acquisition protocols. 1007/S10140-007-0618-9
SCANNER-SPECIFIC CT SIMULATOR 2095

3. Smith-Bindman R, Kwan ML, Marlow EC, et al. Trends in use 22. Syben C, Michen M, Stimpel B, Seitz S, Ploner S, Maier AK.
of medical imaging in US Health Care Systems and in Ontario, Technical note: PYRO-NN: python reconstruction operators in
Canada, 2000–2016. JAMA. 2019;322(9):843-856. doi:10.1001/ neural networks. Med Phys. 2019;46(11):5110-5115. doi:10.1002/
JAMA.2019.11456 mp.13753
4. Kocher KE, Meurer WJ, Fazel R, Scott PA, Krumholz HM, 23. McKenney SE, Nosratieh A, Gelskey D, et al. Experimental vali-
Nallamothu BK. National trends in use of computed tomography dation of a method characterizing bow tie filters in CT scanners
in the emergency department. Ann Emerg Med. 2011;58(5):452- using a real-time dose probe. Med Phys. 2011;38(3):1406-1415.
462.e3. doi:10.1016/J.ANNEMERGMED.2011.05.020 doi:10.1118/1.3551990
5. Mettler J, Wiest PW, Locken JA, Kelsey CA. CT scanning: patterns 24. Hernandez AM, Boone JM. Tungsten anode spectral model using
of use and dose. J Radiol Prot. 2000;20(4):353-359. doi:10.1088/ interpolating cubic splines: unfiltered x-ray spectra from 20 to
0952-4746/20/4/301 640 kV. Med Phys. 2014;41(4):1-15. doi:10.1118/1.4866216
6. Bach PB, Mirkin JN, Oliver TK, et al. Benefits and harms of CT 25. Swinehart DF. The Beer–Lambert law. J Chem Educ.
screening for lung cancer:a systematic review.JAMA—J Am Med 1962;39(7):333-335. doi:10.1021/ed039p333
Assoc. 2012;307(22):2418-2429. doi:10.1001/jama.2012.5521 26. Elam WT, Ravel BD, Sieber JR. A new atomic database for X-ray
7. de Koning HJ, van der Aalst CM, de Jong PA. Reduced spectroscopic calculations. Radiat Phys Chem. 2002;63(2):121-
lung-cancer mortality with volume CT screening in a random- 128. doi:10.1016/S0969-806X(01)00227-4
ized trial. N Engl J Med. 2020;382(6):503-513. doi:10.1056/ 27. Boone JM, Chavez AE. Comparison of x-ray cross sections
NEJMoa1911793 for diagnostic and therapeutic medical physics. Med Phys.
8. VanSonnenberg E, Casola G, Ho M, et al. Difficult thoracic 1996;23(12):1997-2005. doi:10.1118/1.597899
lesions: CT-guided biopsy experience in 150 cases. Radiology. 28. Fujita H, Tsai DY, Itoh T, et al. A simple method for determining
1988;167(2):457-461. doi:10.1148/radiology.167.2.3357956 the modulation transfer function in digital radiography.IEEE Trans
9. Erickson SJ, Hogan QH. CT-guided injection of the stellate Med Imaging. 1992;11(1):34-39. doi:10.1109/42.126908
ganglion: description of technique and efficacy of sympathetic 29. Siewerdsen JH, Antonuk LE, El-Mohri Y, Yorkston J, Huang W,
blockade.Radiology.1993;188(3):707-709.doi:10.1148/radiology. Cunningham IA. Signal, noise power spectrum, and detective
188.3.8351337 quantum efficiency of indirect-detection flat-panel imagers for
10. Kang E, Min J, Ye JC. A deep convolutional neural network using diagnostic radiology. Med Phys. 1998;25(5):614-628. doi:10.1118/
directional wavelets for low-dose x-ray CT reconstruction. Med 1.598243
Phys. 2017;44(10):e360-e375. doi:10.1002/MP.12344 30. Feldkamp LA, Davis LC, Kress JW. Practical cone-beam algo-
11. Adler J, Öktem O. Deep Bayesian inversion. arXiv. 2018. doi:10. rithm. J Opt Soc Am. 1984;1(6):612-619.
48550/arXiv.1811.05910 31. Sun Nuclear. Multi-Energy CT Phantom. Accessed January
12. Kang E, Chang W, Yoo J, Ye JC. Deep convolutional framelet 6, 2022. https://www.sunnuclear.com/uploads/documents/
denosing for low-dose CT via wavelet residual network. IEEE datasheets/Diagnostic/MECT_Phantom_102121.pdf
Trans Med Imaging. 2018;37(6):1358-1369. doi:10.1109/TMI. 32. Kayugawa A, Ohkubo M, Wada S. Accurate determination of ct
2018.2823756 point-spread-function with high precision. J Appl Clin Med Phys.
13. Bhadauria HS, Dewal ML. Medical image denoising 2013;14(4):216-226. doi:10.1120/jacmp.v14i4.3905
using adaptive fusion of curvelet transform and total 33. Kwan ALC, Boone JM, Yang K, Huang SY. Evaluation of the spa-
variation. Comput Electr Eng. 2013;39(5):1451-1460. tial resolution characteristics of a cone-beam breast CT scanner.
doi:10.1016/J.COMPELECENG.2012.04.003 Med Phys. 2007;34(1):275-281. doi:10.1118/1.2400830
14. World Intellectual Property Organization International Publication 34. Balakrishnan V. All about the dirac delta function(?). Resonance.
date: 16 November 2017 International Publication Number: WO 2003;8(8):48-58. doi:10.1007/bf02866759
2017/194787 A1. 35. Beylkin G. Discrete radon transform. IEEE Trans Acoust.
15. Anam C, Sutanto H, Adi K, et al. Development of a computa- 1987;35(2):162-172. doi:10.1109/TASSP.1987.1165108
tional phantom for validation of automated noise measurement 36. De Man B, Basu S, Chandra N, et al. CatSim: a new computer
in CT images. Biomed Phys Eng Express. 2020;6(6):065001. assisted tomography simulation environment. Med Imaging 2007
doi:10.1088/2057-1976/abb2f8 Phys Med Imaging. 2007;6510:65102G. doi:10.1117/12.710713
16. Abadi E, Segars WP, Tsui BMW, et al. Virtual clinical trials in med- 37. Wu M, Fitzgerald P, Zhang J, et al. XCIST—An open access
ical imaging: a review. J Med Imaging. 2020;7(4):042805-1-40. x-ray/CT simulation toolkit. Phys Med Biol. 2022;67(19):194002.
doi:10.1117/1 doi:10.1088/1361-6560/ac9174
17. Segars WP, Bond J, Frush J, et al. Population of anatomically
variable 4D XCAT adult phantoms for imaging research and
optimization. Med Phys. 2013;40(4):1-11. doi:10.1118/1.4794178 S U P P O R T I N G I N F O R M AT I O N
18. Abadi E, Harrawood B, Sharma S, Kapadia A, Segars WP, Samei
Additional supporting information can be found online
E. DukeSim: a realistic, rapid, and scanner-specific simulation
framework in computed tomography. IEEE Trans Med Imaging. in the Supporting Information section at the end of this
2019;38(6):1457-1465. doi:10.1109/TMI.2018.2886530 article.
19. Caballo M, Michielsen K, Fedon C, Sechopoulos I. Towards 4D
dedicated breast CT perfusion imaging of cancer: development
and validation of computer simulated images. Phys Med Biol.
2019;64(24). doi:10.1088/1361-6560/ab55ac How to cite this article: Tunissen SAM,
20. Saunders RS, Samei E. A method for modifying the image Oostveen LJ, Moriakov N, et al. Development,
quality parameters of digital radiographic images. Med Phys.
2003;30(11):3006-3017. doi:10.1118/1.1621870
validation, and simplification of a
21. Moriakov N, Sonke J-J, Teuwen JLIRE. Learned invertible recon- scanner-specific CT simulator. Med Phys.
struction for cone beam CT. arXiv. 2022. doi:10.48550/arXiv.2205. 2024;51:2081–2095.
07358 https://doi.org/10.1002/mp.16679