Deep silicon-based photon counting CT: A first simulation-based study for assessing perceptual benefits across diverse anatomies

Abstract

Objectives:

To assess perceptual benefits provided by the improved spatial resolution and noise performance of deep silicon-based photon-counting CT (Si-PCCT) over conventional energy-integrating CT (ECT) using polychromatic images for various clinical tasks and anatomical regions.

Materials and Methods:

Anthropomorphic, computational models were developed for lungs, liver, inner ear, and head-and-neck (H&N) anatomies. These regions included specific abnormalities such as lesions in the lungs and liver, and calcified plaques in the carotid arteries. The anatomical models were imaged using a scanner-specific CT simulation platform (DukeSim) modeling a Si-PCCT prototype and a conventional ECT system at matched dose levels. The simulated polychromatic projections were reconstructed with a matched in-plane resolution using manufacturer-specific software. The reconstructed pairs of images were scored by radiologists to gauge task-specific perceptual benefits provided by the Si-PCCT compared to ECT based on visualization of anatomical and image quality features. The scores were standardized as z-scores for minimizing inter-observer variability and compared between the systems for evidence of statistically significant improvement (one-sided Wilcoxon rank-sum test with a significance level of 0.05) in perceptual performance for Si-PCCT.

Results:

Si-PCCT offered favorable image quality and improved visualization capabilities, leading to mean improvements in task-specific perceptual performance over ECT for most tasks. The improvements for Si-PCCT were statistically significant for the visualization of lung lesions (0.08 ± 0.89 vs. 0.90 ± 0.48), liver lesions (−0.64 ± 0.37 vs. 0.95 ± 0.55), soft tissue structures (−0.47 ± 0.90 vs. 0.33 ± 1.24), and inner ear cochlea (−0.47 ± 0.80 vs. 0.38 ± 0.62).

Conclusions:

Si-PCCT exhibited mean improvements in task-specific perceptual performance over ECT for most clinical tasks considered in this study, with statistically significant improvement for 6/20 tasks. The perceptual performance of Si-PCCT is expected to improve further with availability of spectral information and reconstruction kernels optimized for high resolution provided by smaller pixel sizes of Si-PCCT. The outcomes of this study indicate the positive potential of Si-PCCT for benefiting routine clinical practice through improved image quality and visualization capabilities.

INTRODUCTION:

Photon-counting computed tomography (PCCT) is currently driving a revolution in clinical diagnostic imaging by virtue of its ability to provide improved spatial resolution, energy weighting leading to optimal signal-to-noise and contrast-to-noise ratios, and material-specific imaging capabilities¹. An emerging PCCT detector technology is the deep silicon-based detectors. These detectors have a novel edge-on design utilizing thin but deep silicon wafers interfolded with tungsten strips^2–4. This detector design is currently receiving interest from the imaging community due to its potential for providing enhanced spatial and spectral resolution, improved noise performance, and reduction of the adverse effects of pulse pileup.

Due to this unique design that promises improved technical performance, pilot studies are already underway^5,6 to evaluate a Si-PCCT prototype for imaging capabilities and optimized system design. Of all the potential benefits of Si-PCCT, enhanced perceptual performance has potentially an immediate impact on clinical imaging due to the system’s improved spatial resolution and noise performance.

Although there are multiple prior studies that demonstrate how CdTe/CZT-PCCT systems improve spatial resolution and noise performance^7–12, such studies have been rare for Si-PCCT prototypes. One example is the study by Silva et al.¹³ which demonstrated an increase in spatial resolution over a commercial state-of-the-art CT scanner using a resolution phantom. This finding was made more evident using images of a skull phantom. The resolution improvements offered by Si-PCCT have been further demonstrated by a simulation study from Sundberg et al.¹⁴ utilizing shape measurements of charge clouds from Compton interactions in deep silicon detectors. Although these studies have been beneficial for establishing the baseline imaging performance of Si-PCCT, they lack task-based evaluations that are needed to effectively translate Si-PCCT technology to routine clinical practice. A few studies have attempted to perform such evaluations for Si-PCCT for clinical tasks that focus on stent appearance¹⁵, quantification of radiomics features¹⁶, and pediatric imaging¹⁷, but they are still far from comprehensive.

These evaluations are conventionally done using scans from human subjects. These imaging experiments are challenging to perform because they are expensive, slow, and have ethical limitations due to the requirement of exposing human subjects to ionizing radiation. In addition, such studies are especially impractical for Si-PCCT due to the limited access to prototype scanners. These limitations can be overcome through simulation-based imaging studies, often referred to as virtual imaging trials¹⁸, where imaging experiments are emulated using accurate models of human subjects and imaging systems.

The purpose of this article was to use the novel methodology of virtual imaging to assess the task-specific perceptual benefits of Si-PCCT over ECT using polychromatic images for a range of clinical tasks and anatomical regions (liver, lungs, inner ear, and head-and-neck).

MATERIALS AND METHODS:

Computational human models

A human model (XCAT)¹⁹ was selected representing the 50^th percentile BMI of the adult male population in the United States. This human model was used as the basis for developing the various anatomies and abnormalities considered in this study. The selected XCAT model was voxelized at an isotropic resolution of 0.1 mm. Materials inside XCAT were defined with densities and elemental compositions from recommendations of the International Commission on Radiation Units and Measurements (ICRU)²⁰.

The XCAT lungs were enhanced with a detailed lung architecture including airways, vasculature, and the parenchymal structures using an iterative anatomically-based volume-filling branching algorithm^21,22. For the liver, a detailed vasculature tree was incorporated using an anatomically-based algorithm based on diffusion-limited aggregation²³. The lungs and liver included lesions that were generated using a dynamic nutrient-access-based stochastic model^24,25. The lesions were manually inserted into the parenchymal regions of the respective organs ensuring no intersections with major anatomical structures. The densities and material composition of the lesions were modeled as that of soft tissue. To simulate contrast enhancement typical of liver imaging, elemental compositions of materials in XCAT were modified to incorporate iodine concentrations corresponding to the portal venous phase²⁶.

XCAT was also modified to enable head-and-neck (H&N) imaging conditions. To do so, H&N plaque models²⁷ were generated from segmentations of histological images displaying calcification with a lipid cap and inserted into the carotid arteries. To enable contrast-enhanced H&N imaging simulations, densities of materials in the XCAT were modified to achieve target HU values representative of the arterial phase²⁸.

XCAT was further conditioned for imaging experiments of the inner ear. The inner ear was modeled using segmentations of a 0.2 mm high-resolution imaging dataset of a physical head phantom consisting of a human skull embedded in soft tissue material, which was then upsampled to 0.1 mm to match the voxel size of the selected XCAT phantom. The soft tissue material present in the nasal and mastoid air cavities of the head phantom was substituted with air during the segmentation process. Anatomical structures that were not conspicuous in the imaging dataset of the head phantom, such as outer auditory canals, eardrum, cochlea, and the stapes, malleus, and incus, were augmented with corresponding structures from the original XCAT phantom.

Simulation pipeline for CT

The modeled phantoms were imaged using a CT simulation platform (DukeSim²⁹), modeling the scanner-specific geometry and components of a representative Si-PCCT prototype (GE HealthCare, Waukesha, WI) and a conventional clinical ECT system (Revolution^™, GE Healthcare, Waukesha, WI). For both ECT and Si-PCCT systems, a helical scan was simulated for all anatomies with clinically representative acquisition settings as summarized in Table 1.

Table 1.

A summary of the scan protocols utilized for simulating helical scans for the different anatomical regions considered in this study.

Anatomy	Tube voltage (kV)	Tube current (mA)	Rotation time (s)	Collimation (mm)	Pitch	CTDIvol (mGy)
Lung	120	417	0.5	38.4	0.98	12.95
Liver	120	500	0.5	38.4	0.98	15.52
Inner Ear	120	208	0.5	38.4	0.98	6.46
Head-and-Neck (H&N)	120	750	0.5	38.4	0.98	46.57

The Si-PCCT prototype considered in this study uses the same x-ray tube and scanner geometry as Revolution APEX^™ GE HealthCare, Waukesha, WI). The prototype utilizes silicon-based edge-on-irradiated sensors^3,13 with 8 energy bins implemented using variable energy thresholds³⁰. The simulation of the detection process involved binning of photon counts into 28 energy bins (0–140 keV, bin width: 5 keV) to simulate an “ideal” photon-counting detector, post-processing the photon counts in energy bins to incorporate realistic detection efficiency and spatio-energetic crosstalk characteristics of a deep silicon photon-counting detector, and addition of Poisson-distributed quantum noise to all energy bins. The detection efficiency refers to the signal conversion efficiency of the sensors quantified as the number of photons detected per incident photon on the detector for a given energy bin. The spatio-energetic crosstalk characteristics refer to the spatial and energetic correlations introduced between neighboring pixels due to x-ray crosstalk and charge sharing, which further affect the photon counts in energy bins across all pixels in a pixel neighborhood. For this study, the factors modeling these effects were derived from previously published detector models³¹. The effect of pulse pileup on the response of the detector was not modeled in this study, which is a reasonable assumption for Si-PCCT^3,32.

To provide a conservative estimate of differences in performance, the Si-PCCT system design was chosen to represent the worst-case scenario for spatial resolution, with maximum detector pixel sizes chosen for both x and z directions from a range of pixel sizes being considered for system development (physical pixel sizes: 0.2–0.4 mm in x, 0.5–0.7 mm in z). Due to the minor contribution of object scatter to the total signal for the chosen kV and collimation settings used in this study³³, scatter was not modeled in the imaging simulations. The focal spot was modeled as a rectangular uniform intensity emission source with a size (0.4 mm x 0.75mm) selected to match the small pixel size of Si-PCCT while still supporting the dose levels considered in this study. For modeling automatic exposure control, the tube current per projection was modulated based on water-equivalent diameters estimated from simulated anterior-posterior and left lateral localizers³⁴. For generating the Si-PCCT sinograms for reconstruction, the simulated photon counts in all energy bins were summed without weighting (i.e., using equal weights of 1 for all energy bins) to obtain polychromatic sinograms and then corrected for the effects of beam hardening using a 5^th order polynomial water correction³⁵. The coefficients of the polynomial used for corrections were estimated by fitting the polynomial to monoenergetic simulations of detector signal attenuated with different thicknesses of water.

The ECT system used for comparison was Revolution APEX^™ (GE HealthCare, Waukesha, WI) and was modeled using the same simulation platform with a detector pixel size of 1.1 mm in both x and z. The photon counts were similarly post-processed to incorporate the detection efficiency and spatial crosstalk characteristics of an energy-integrating detector, followed by the addition of Poisson-distributed quantum noise, and weighting with respective bin energies before summing to generate polychromatic sinograms. Similar to Si-PCCT, the focal spot for the ECT system was modeled as a rectangular uniform emission intensity source with dimensions corresponding to the smallest available focal spot on the system.

Simulated sinograms for both systems were reconstructed using a manufacturer-specific filtered-backprojection reconstruction software (GE Healthcare, Waukesha, WI) with settings representative of routine clinical practice as summarized in Table 2. Although analytical algorithms such as the filtered-backprojection are becoming increasingly obsolete for CT reconstructions³⁶, they are ideal for this study as they minimize the confounding effect of the choice of the reconstruction algorithm and allow for a fairer assessment of the impact of detector technology on image quality. To assess the perceptual benefits provided by Si-PCCT over ECT for existing clinical practice, the in-plane resolution of Si-PCCT images was matched to ECT by using MTF-matching reconstruction kernels for soft-tissue imaging (“stnd”, MTF10: 6.1 lp/cm) and high-resolution imaging (“bone”, MTF10: 10.0 lp/cm). The benefits from improved z-resolution of the Si-PCCT were allowed for evaluation by reconstructing the images for both systems at native slice thicknesses.

Table 2.

A summary of the parameters utilized for reconstructing sinograms of simulated helical scans for the different anatomical regions considered in this study.

Anatomy	FOV (cm)	Matrix Size (pixels)	Kernel	Slice thickness (Si-PCCT, ECT) (mm)
Lung	15	512×512	stnd	0.4, 0.6
Liver	30	512×512	stnd	0.4, 0.6
Inner Ear	10	512×512	bone	0.4, 0.6
Head-and-Neck (H&N)	15	512×512	stnd	0.4, 0.6

Image evaluation

For assessing perceptual differences in images for Si-PCCT vs. ECT, differences in rendering of clinically relevant anatomical features and image quality characteristics were evaluated by a cohort of six experienced radiologists from a diversity of specialties (abdominal, cardiothoracic, and musculoskeletal imaging). Each participating radiologist was asked to perform a blinded A vs. B comparison of images simulated for Si-PCCT and ECT and score the images on a scale of 1–100 (least favorable-most favorable: 1–100) for prompts designed to highlight specific anatomical and image quality features that are relevant to the anatomy of interest, as summarized in Table 3.

Table 3.

A summary of the prompts utilized by observers for scoring simulated images of the different anatomical regions considered in this study.

Anatomy	Prompts
Lung	Visualization of the lesion Visualization of lung parenchyma Visualization of soft tissue-rib interface Favorability of texture in lung region Favorability of image artifacts
Liver	Visualization of lesion Visualization of vessels and organ boundaries Favorability of texture in liver region Favorability of image artifacts
Inner Ear	Visualization of inner ear bones (malleus, incus, stapes) Visualization of soft tissue structures (such as the eardrum/ tympanic membrane) Visualization of the cochlea Visualization of the mastoids and the associated air cavities Favorability of texture in soft tissue regions Favorability of image artifacts
Head-and-Neck (H&N)	Visualization of plaque Visualization of the separation between plaque and lumen Resolution of vessel boundaries Differentiation of soft and calcified plaques Favorability of texture in soft tissue regions

For all anatomies, paired stacks of simulated images in sagittal and axial views for both Si-PCCT and ECT systems were shown to the observers using a clinical image display application (Advantage Workstation Server 3.2, GE Healthcare, Waukesha, WI). For the axial view, the slice thicknesses for image display between the systems were closely matched (Si-PCCT: 1.20 mm, ECT: 1.25 mm), with an option to scale to native slice thicknesses as needed. For the sagittal view, the pair of images were displayed at native slice thickness (Si-PCCT: 0.4 mm, ECT: 0.6 mm). The default window widths and levels (W/L) were selected to highlight the differences between the pair of images, with changes allowed as needed during the image review.

To minimize inter-observer variability, scores assigned by a given radiologist for a prompt were converted to standard scores (z) using mean and standard deviation estimated from all scores assigned by that radiologist for the related anatomy. To statistically compare the standard scores assigned to ECT and Si-PCCT systems, a one-sided Wilcoxon rank-sum test with a significance level of 0.05 was performed using statistical computing software (R statistical package v.4.0.3, The R Foundation, Vienna, Austria).

RESULTS:

Figures 1–4 show simulated images for Si-PCCT (left) and ECT (right) for the different anatomical regions considered in this study. It was observed that the visualization capabilities and image quality offered by Si-PCCT were comparable to ECT when comparing images with matched in-plane resolution, with discernible improvements for certain features as described below.

Figure 1.

Simulated sagittal images (W/L: 1200/−600) of lungs with a lesion using an energy-integrating CT (ECT) (slice thickness: 0.6 mm) (left) and a deep silicon-based photon-counting CT (Si-PCCT) (slice thickness: 0.4 mm) (right), reconstructed using a “stnd” kernel. The corresponding noise measurements for a region-of-interest (ROI) placed within the lung parenchyma are 19.2 HU and 21.8 HU for ECT and Si-PCCT, respectively. Compared to ECT, Si-PCCT images exhibited improved visualization of lung parenchyma structures, especially those with fine details such as secondary pulmonary lobules and the lesion boundary. The location of the lesion is marked with an arrow in both images.

Figure 4.

Simulated coronal images (W/L: 700/200) of head-and-neck (H&N) with plaques in carotid arteries using an energy-integrating CT (ECT) (slice thickness: 0.6 mm) (left) and a deep silicon-based photon-counting CT (Si-PCCT) (slice thickness: 0.4 mm) (right), reconstructed using the “stnd” kernel. The corresponding noise measurements for a region-of-interest (ROI) placed within the soft tissue region are 5.0 HU and 5.5 HU for ECT and Si-PCCT, respectively. Compared to ECT, the Si-PCCT images offered enhanced visualization of both soft and calcified plaques, including improved definition of plaque boundaries. The locations of the plaques are marked with arrows in both images.

For lung, Si-PCCT images exhibited improved visualization of lung parenchyma structures with increased conspicuity and sharpness, especially those with fine details such as secondary pulmonary lobules and lesion boundaries. The enhancements in the visualization of lesion morphology were observed to be most evident in the sagittal view, as shown in Figure 1, with clarity of lesion spiculations being higher for Si-PCCT compared to ECT. For liver, as shown in Figure 2, Si-PCCT offered favorable noise texture with finer graininess and improved low-contrast visualization with enhanced conspicuity of lesion boundaries and the portal vein. For inner ear, as shown in Figure 3, Si-PCCT offered improved visualization of fine anatomical structures such as ossicles of the inner ear, mastoids, and the associated mastoidal air cavities. Improvements in the visualization of soft tissue structures of the inner ear such as the tympanic membrane were also observed in the images for Si-PCCT. For H&N, as shown in Figure 4, Si-PCCT images exhibited enhanced visualization of both soft and calcified plaques, characterized by sharper definition of plaque boundaries due to increased calcium contrast and reduced blooming.

Figure 2.

Simulated axial images (W/L: 400/40) of the liver with lesion using an energy-integrating CT (ECT) (slice thickness: 1.25 mm) (left) and a deep silicon-based photon-counting CT (Si-PCCT) (slice thickness: 1.2 mm) (right), reconstructed using the “stnd” kernel. The corresponding noise measurements for a region-of-interest (ROI) placed within the liver parenchyma are 18.7 HU and 19.5 HU for ECT and Si-PCCT, respectively. Compared to ECT, the Si-PCCT images offered favorable noise texture with finer graininess, improved low-contrast visualization, and enhanced conspicuity of lesion boundaries and the portal vein. The location of the lesion is marked with an arrow in both images.

Figure 3.

Simulated axial images (W/L: 2000/350) of the inner ear using an energy-integrating CT (ECT) (slice thickness: 0.6 mm) (top left) and a deep silicon-based photon-counting CT (Si-PCCT) (slice thickness: 0.4 mm) (top right), reconstructed using the “bone” kernel. The image in the bottom row shows the corresponding high-dose (2x dose) image for Si-PCCT. The corresponding noise measurements for a region-of-interest (ROI) placed within the soft tissue region are 80.3 HU and 68.1 HU for ECT and Si-PCCT, respectively. The image insets show magnified regions of the image exhibiting notable differences in visualization between ECT and Si-PCCT. Compared to ECT, the Si-PCCT images provided improved visualization of finer anatomical structures such as ossicles of the inner ear, mastoids, and the tympanic membrane. The location of structures showing notable improvements in visualization are marked with arrows in both images.

The observer study also demonstrated noticeable improvements in the visualization performance and image quality for Si-PCCT over ECT, as shown in Figure 5. The mean observer scores were noted to be higher for Si-PCCT over ECT for all features across all anatomical regions considered in this study, with the exception of favorability of artifacts in lungs (−0.19 vs. −0.32) and differentiation between soft and calcified plaques (0.61 vs. 0.50), for which the mean scores were lower. The Wilcoxon rank-sum test revealed statistically significant improvements in scores (p <= 0.05) for Si-PCCT over ECT for the following features: visualization of lesion (0.08 ± 0.89 vs. 0.90 ± 0.48) in lungs; visualization of lesion (−0.64 ± 0.37 vs. 0.95 ± 0.55) and favorability for texture (−1.07 ± 0.81 vs. 0.45 ± 0.48) in liver; visualization of soft tissue structures (−0.47 ± 0.90 vs. 0.33 ± 1.24) and cochlea (−0.47 ± 0.80 vs. 0.38 ± 0.62), and favorability for texture (−0.41 ± 0.71 vs. 0.36 ± 0.86) in inner ear.

Figure 5.

Standard scores (z) for all prompts across the different anatomies (lungs, liver, inner ear, and head-and-neck (H&N)) considered in this study for an energy-integrating CT (ECT) (red) and a deep silicon-based photon-counting CT (Si-PCCT) (blue). For all prompts, higher preference and favorability for visualization capabilities and image quality features receive a higher score. The circles and the horizontal bars indicate the mean and standard deviation of standard scores, respectively. The prompts for which the scores were statistically significantly higher for Si-PCCT compared to ECT have been highlighted with an asterisk (*).

DISCUSSION:

This article reported the outcomes of the first comprehensive virtual imaging study assessing the task-specific perceptual benefits provided by Si-PCCT over ECT across a range of diverse anatomies. As indicated by the visual differences in simulated images and the associated observer scores from a cohort of experienced radiologists, Si-PCCT offered improved image quality and enhanced visualization of clinically relevant features, resulting in improved perceptual performance over ECT. Due to the matched in-plane resolution of both systems, the improvements in visualization performance resulting from improved spatial resolution of Si-PCCT were found to be most dominant in the sagittal plane due to thinner native slice thickness (0.4 mm) compared to ECT (0.6 mm). For the axial plane, the improvements in visual performance for Si-PCCT over ECT were primarily to improvements in image noise from equal energy weighting of photons³⁷ and reduced noise aliasing^38,39, given the effects of electronic noise mitigation are not expected to be dominant at dose levels considered in this study. The improvements in visualization performance for Si-PCCT over ECT can also be attributed to improvements in image contrast resulting from equal weighting for all photon energies⁴⁰, especially for regions containing high contrast materials such as bone or contrast agents.

The visual improvements observed in images for Si-PCCT over ECT also translated to higher task-specific perceptual performance in the clinic, with improvements in mean observer scores for most tasks and statistically significant improvements for 6/20 tasks that include the visualization of lesions in the lung and liver and specific features of the inner ear such as the cochlea and the tympanic membrane. The demonstrated improvements in the perceptual performance of Si-PCCT over ECT across a diversity of clinical tasks strongly indicate the relevance of this upcoming technology for having an immediate impact on current clinical practice. Furthermore, with the availability of spectral information that can be utilized for generating energy-weighted images that maximize contrast-to-noise ratio for specific features⁴¹ and reconstruction kernels that are optimized for reducing noise while preserving native spatial resolution resulting from smaller pixel sizes of Si-PCCT⁴², the perceptual performance of Si-PCCT could be expected to improve further.

This study had some limitations. First, the simulation platform (DukeSim) was not validated against the exact scanners modeled in this study. However, DukeSim has been validated in the past for numerous other clinical scanners^43–46, which provide robust evidence of its ability to simulate images appropriate for clinical evaluations given the availability of realistic input models describing the image formation process. Despite this limitation, the use of simulations for imaging evaluations has the unique benefit of enabling paired comparisons without patient-related confounders, which is practically impossible to achieve in traditional imaging studies due to the hazards of exposing human subjects to ionizing radiation. Second, the evaluation of the perceptual benefits was performed using polychromatic images without considering binned spectral data and high-resolution information provided by small pixel size of Si-PCCT. The utilization of high-resolution information for visualizing finer details and spectral data (both through the generation of virtual monoenergetic images and material maps) has great utility for maximizing the task-specific perceptual benefits of Si-PCCT and needs to be considered in future studies. Third, for lung imaging tasks, the impact of respiratory and cardiac motion was not studied, which could potentially lead to the overestimation of perceptual benefits provided by Si-PCCT over ECT. Fourth, although virtual imaging studies provide an effective avenue for conducting task-specific evaluations of upcoming imaging technologies, they are limited in terms of their realism, inhibiting their capability to fully replace traditional studies utilizing human subjects and physical scanners. However, these virtual studies in their current form still have immense potential to aid the design and optimization of new technologies while also serving as precursors to traditional trials for more targeted investigations.

This article reported the outcomes of the first comprehensive virtual imaging study assessing the task-specific perceptual benefits from improved spatial resolution and noise performance of Si-PCCT over ECT. For all anatomical regions considered in the study, Si-PCCT exhibited enhanced image quality over ECT, with apparent qualitative improvements in visualization of relevant features, leading to enhanced perceptual performance as indicated by improvements in mean observer scores for most clinical tasks. The improvements in task-specific perceptual performance for Si-PCCT were found to be statistically significant for visualization of lesions in the lung and liver and features of the inner ear such as the cochlea and the tympanic membrane. The outcomes of this study indicate the positive potential of this technology for improving current clinical practice.

Abbreviations and Acronyms:

CT

computed tomography

Si-PCCT

deep silicon-based photon-counting computed tomography

Si

PCD deep Si photon-counting detector

ECT

energy-integrating computed tomography

XCAT

extended cardiac-torso

H&N

head-and-neck