Development of physiologically-informed computational coronary artery plaques for use in virtual imaging trials

Abstract

Background:

As a leading cause of death, worldwide, cardiovascular disease is of great clinical importance. Among cardiovascular diseases, coronary artery disease (CAD) is a key contributor, and it is the attributed cause of death for 10% of all deaths annually. The prevalence of CAD is commensurate with the rise in new medical imaging technologies intended to aid in its diagnosis and treatment. The necessary clinical trials required to validate and optimize these technologies require a large cohort of carefully controlled patients, considerable time to complete, and can be prohibitively expensive. A safer, faster, less expensive alternative is using virtual imaging trials (VITs), utilizing virtual patients or phantoms combined with accurate computer models of imaging devices.

Purpose:

In this work, we develop realistic, physiologically-informed models for coronary plaques for application in cardiac imaging VITs.

Methods:

Histology images of plaques at micron-level resolution were used to train a deep convolutional generative adversarial network (DC-GAN) to create a library of anatomically variable plaque models with clinical anatomical realism. The stability of each plaque was evaluated by finite element analysis (FEA) in which plaque components and vessels were meshed as volumes, modeled as specialized tissues, and subjected to the range of normal coronary blood pressures. To demonstrate the utility of the plaque models, we combined them with the whole-body XCAT computational phantom to perform initial simulations comparing standard energy-integrating detector (EID) CT with photon-counting detector (PCD) CT.

Results:

Our results show the network is capable of generating realistic, anatomically variable plaques. Our simulation results provide an initial demonstration of the utility of the generated plaque models as targets to compare different imaging devices.

Conclusions:

Vast, realistic, and variable CAD pathologies can be generated to incorporate into computational phantoms for VITs. There they can serve as a known truth from which to optimize and evaluate cardiac imaging technologies quantitatively.

I. Introduction

Cardiovascular disease has consistently been a leading cause of mortality among men and women worldwide for decades, accounting for nearly 30% of all deaths. The most prevalent type of cardiovascular disease is coronary artery disease (CAD), accounting for roughly 10% of all deaths. For meaningful diagnosis and accurate evaluation of the state of any disease, diagnostic imaging is critically important—and a widely available modality with specialized acquisition protocols for noninvasive CAD applications is CT. Noninvasive diagnostic imaging of CAD is difficult due to the size scale of the disease relative to the magnitude of cardiac motion—which is unavoidable in vivo. Precedent suggests that validation and optimization of clinical diagnostic technology is best evaluated in a clinical trial. However, the current pace of medical and technological advances, coupled with ever-present concerns about the safety, practicality, and ethics of optimizing ionizing imaging protocols on human subjects necessitates a robust alternative: the virtual imaging trial.

Virtual imaging trials are computational studies that use anthropomorphic models of their targets in tandem with physics-based imaging simulations—in this case, CT. The results of these imaging simulations can be analyzed to evaluate the effects of image acquisition and reconstruction parameters on task-specific image quality. Details of the models themselves (e.g., normal anatomy, pathology) can be modified systematically to ascertain the performance of the imaging protocol as a function of these changes.

Prior work in simulating CAD has focused on coronary or carotid arterial stenosis, or plaques evaluated based on image-based ground truth^1–4 using mathematical models of idealized situations which answer general questions (e.g., changes in blood pressure across the spatial extent of a plaque)¹, or finite element analysis (FEA) simulations intended to predict the conditions under which a plaque may be more likely to form or may be at greater risk of incurring an adverse event in both single vessels and networks of vessels^5–13. However, this does not provide a model sufficient to simulate the preferred clinical evaluation technique of coronary CT angiography (CCTA). This limitation restricts virtual imaging simulation studies intended to optimize the parameters that determine the quality of the final image and hence the quality of the diagnostic information contained.

Virtual imaging trials require, at minimum, a realistic virtual patient and a platform for simulating a clinically-relevant imaging system. Significant work has been done in this laboratory to advance both the virtual patient, through the development of the series of XCAT phantoms^14–30 modeling various patient anatomies, as well as the simulation platform, through the development of DukeSim³¹ capable of accurately modeling modern imaging systems. Many of the basic anatomical details of a given XCAT phantom come from CT image segmentation. This poses a challenge for the state of a virtual representation of CAD due to the conflation of the effects of cardiac motion blurring, compositional variability of cardiac plaques, and the size of the disease relative to the spatial resolution of CT ^32,33.

This work presents a method for using plaque histology and morphology data acquired at micron-level resolution by Elucid Bioimaging^3,34 to generate new, novel plaques informed by a large, original patient cohort. Methodology for curating and validating synthesized plaques’ anatomical and physiological realism is also incorporated to ensure quality. The plaque models created here can be incorporated into the cardiac models of the XCAT phantoms to simulate various states of CAD for virtual imaging trials.

II. Materials and Methods

Overview

Several sets of segmented, high-resolution histology images of atherosclerotic plaques were used to train a deep learning network to generate anatomically variable plaque images. Each generated plaque image was post-processed using a segmentation algorithm to eliminate floating-point values, resulting in unique integer values designating the structures of the plaque and surrounding vessel. The adventitia layer of the vessel, not segmented in the histological data, was added by dilating the image to a defined thickness value. The 2D image model was then extruded about the third dimension to a user-defined length to produce a 3D voxel representation. The stability of each plaque was assessed by converting it into a volumetric mesh and evaluating it by finite element analysis (FEA). These methods were utilized to create a library of anatomically variable plaque models. Example plaques were incorporated into the beating heart model of the 4D XCAT computational phantom to perform initial simulations comparing standard energy-integrating detector (EID) CT with photon-counting detector (PCD) CT. More details on these methods and pilot study can be found in the following sections.

Atherosclerotic plaque data cohort

The plaque data in this study consisted of radiologist- and pathologist-annotated diseased arteries, which were excised and evaluated histologically (under ClinicalTrials.gov Identifier: NCT02143102). For this study, the utilized dataset consisted of histologically-segmented plaques obtained from excised carotid arteries. The plaques were excised from patients and processed in a zero-stress configuration (i.e., with the lumen of the affected artery open). They were processed such that they produced an approximately in vivo slice of the atherosclerotic plaque.

The atherosclerotic artery images were two-dimensional RGB maps with colors corresponding to different tissue types, including vessel lumen and external tissue in addition to the pathologies: intra-plaque hemorrhage (IPH) present in 38% of segmentations, dense calcification (CALC) present in 56% of segmentations, lipid-rich necrotic core (LRNC) present in 93% of segmentations, and matrix (fibrous) tissue (MATX) present in 100% of segmentations with pathology present (see ³⁵ for characterization of objective, histologically-defined tissue type by CTA). An illustration of the approximate morphology and location of the pathologies associated with this data set are shown in Fig. 1, and more detailed statistics on the composition of the plaques are shown in Fig. 2.

Figure 1.

Diagram of a generic atherosclerotic plaque to illustrate approximate morphologies and relative position of individual types of pathologies within the coronary artery. The pathologies shown here (in order of decreasing prevalence within the data set) are fibrous matrix (MATX), lipid-rich necrotic core (LRNC), calcification (CALC), and intra-plaque hemorrhage (IPH).

Figure 2.

(a): Bar chart indicating the number of plaques in the cohort with given material percent-composition in 10-percentile bins (e.g., there are roughly 200 plaques consisting of 90–100% fibrous matrix, etc.). (b): Percent-composition of each plaque by tissue type for each of the 756 plaques in the cohort. (b): may also be thought of macroscopically (i.e., percent-composition of the entire cohort for any tissue type has a colored region of the correct relative area.)

The prevalence of pathologies in terms of total segmented pathology pixels (i.e., excluding lumen and external pixels) was 2%, 8%, 17%, and 73% for IPH, CALC, LRNC, MATX respectively—indicating that the cases analyzed here are representative of a diverse and extensively diseased cohort. Only pathologies were included in the segmentations used here (i.e., from the diseased media and radially inward toward the lumen). Each plaque and its surrounding (empty) space had a size of 128 by 128 pixels. The plaque segmentations numbered 756 in total, each originating from a unique plaque section.

Generation of novel atherosclerotic plaques informed by the original cohort

To increase the cohort size and anatomical variability of the data set, a deep learning method was employed to create a model of the anatomical and pathological features of the segmented plaque cohort. This model allowed for the generation of novel, synthetic atherosclerotic plaques with realistic composition and spatial distribution to be used in coronary CT angiography (CCTA) simulation studies within the XCAT–DukeSim^{28–30,36–38} framework.

The plaque segmentations were two-dimensional arrays (images) of similar physical size with properties that allowed for fast data augmentation (Fig. 3). This data was used to train a deep convolutional generative adversarial network (DC-GAN)^39,40 that included a generator and a discriminator (Fig. 4). The generator network synthesized a distribution of images intended to mimic the qualities of the original plaques visually. The discriminator network informed the parameter optimization of the generator model by minimizing the difference between the synthetic and real distributions of images. DC-GAN was chosen due to its stability and efficiency in generating high-resolution images, while still maintaining a relatively simple architecture to facilitate training and implementation.

Figure 3.

Illustration of the data augmentation process using a single plaque image. Individual plaques are rotated arbitrarily about the center of their bounding box. Each of the rotated plaques is additionally mirrored (about any plane through the lumen), resulting in multiple realizations of the same plaque in different two-dimensional orientatins. This strategy was used to increase the size of the training set while avoiding overfitting the algorithm to specific individual pathology locations while still making use of each original training image.

Figure 4.

GAN schematic illustrating the general principle of the iterative training process for the generator. Step A initializes the generator model with the random vector, which the generator uses to produce noise in step B. In step C, the synthetic (generated) images are randomized among the original segmentation images with the ground-truth withheld until Step F. Steps D and E indicate that the training and synthetic data are combined as input into the discriminator network. In step F, the ground truth labels of synthetic and training data are used to determine the ability of the discriminator network in discriminating between the two classes of images. In step F, this information is fed back into the generator. Steps B–G repeat once per training epoch.

The DC-GAN model was implemented in PyTorch⁴¹. In addition to data augmentation, the following model design elements were used to help stabilize training. The generator network consisted of 5 deconvolutional layers. The first deconvolutional layer was size $256\times 512\times 8$, the second $512\times 256\times 4$, the third $256\times 128\times 4$, the fourth $128\times 64\times 4$, and the fifth $64\times 1\times 4$. Each deconvolutional layer in the generator was connected via a leaky rectified linear unit (ReLU), with the final output connected by a hyperbolic tangent. The fifth deconvolutional layer outputs a set of images. The images output by the generator were then used as the input to the discriminator network, which consisted of 5 convolutional layers. The first convolutional layer was size $1\times 64\times 4$, the second $64\times 128\times 4$, the third $128\times 256\times 4$, the fourth $256\times 512\times 4$, and the fifth $512\times 1\times 8$. Each convolutional layer in the discriminator was connected via ReLU. The DC-GAN architecture was trained with a batch size of 32 images for 1000 epochs with a learning rate of $5\times {10}^{-5}$ using the ADAM optimizer with parameters ${\beta }_1=0.5$00 and ${\beta }_2=0.999$⁴². The $L_2$ distance was used as the loss function. It was calculated as the square root of the sum of the squared differences between the pixel values of generated images and the reference images.

The DC-GAN output are grayscale images of size $128\times 128$ pixels with floating-point pixel values distributed as shown in Fig. 5. The output values corresponded relatively closely (visually) to the distribution of pixel values from the original data set. To ensure that the histograms for the training and output data sets matched more closely, the DC-GAN output data set was segmented using an image segmentation method as proposed by Aja-Fernández et al.⁴³. Here, we applied the segmentation method to the output DC-GAN images over five iterations of fuzzy aggregation with a Gaussian smoothing kernel having a standard deviation of two pixels, and a target of six output classes (external, IPH, MATX, CALC, LRNC, and vessel lumen(s)). Histograms comparing the pixel values in the segmented DC-GAN output to the original plaque cohort are overlaid in Fig. 6, demonstrating improved (cf., Fig. 5 for pre-segmentation histogram; Fig. 7 for direct image comparison) correspondence.

Figure 5.

Histogram of pixel values present in the original plaque segmentations (magenta) and in the (continuous data) DC-GAN output (blue).

Figure 6.

Histogram of pixel values present in the original plaque segmentations (magenta) and in the (discrete) iterative fuzzy aggregation segmentation of the DC-GAN output (blue).

Figure 7.

(a): Comparison of 200 raw DC-GAN output images (continuous, floating-point data) to (b), their iterative aggregation segmentation counterparts (discrete, integer data). Very little difference can be seen between the images of the left and right demonstrating that the segmentation algorithm used to convert the floating-point output from the DC-GAN to an integer format, in which integers represent the individual plaque components, does not lose any anatomical details.

The segmented DC-GAN images were modified with a dilation operation to add the adventitia layer to the blood vessel structure as a separate tissue class. This layer was not defined in the original segmented histological data of the plaques and was thus added by using image dilation to add a layer of pixels to the outer boundary of the vessel corresponding to a thickness value of 0.75 mm, an average value for the coronaries⁴⁴, Fig. 8. Additionally, there was one non-tissue integer class; lumen-adjacent pixels were designated as part of the pressure loading surface (intima). The 2D models were then extruded about the third dimension to give them a user-defined length, producing 3D voxel representations. For finite element analysis (following section), an arbitrary length of 2.5 mm was used as it was adequate to perform this analysis. For simulation purposes, the plaques can be made shorter or longer depending on the user’s preference.

Figure 8.

Addition of the adventitia layer (red) to a plaque image. Dilation is used to add multiple layers of pixels to the outer blood vessel corresponding to a thickness of 0.75 mm.

The individual components of the voxelized plaques were used to define isosurfaces for each tissue class. The isosurfaces were used as the input to CGALmesh⁴⁵, which produced multi-class tetrahedral meshes with a radius bound of 200 and a distance bound of 0.5. Each tetrahedral mesh output was then input to Tetgen⁴⁶ for consistency checking. The final tetrahedral mesh components were saved as .node, .ele, .face, and .edge files describing each plaque.

Finite element analysis

Generated output plaques were first evaluated for physiological feasibility (e.g., not having a tissue composition consistent with sudden coronary death as indicated in Kragel et al.¹⁶ and Dollar et al.²² for each plaque’s approximate stenosis level). The remaining (viable) plaques were evaluated via finite element (FE) analysis⁹. Specifically, this analysis focused on internal stress distributions within the diseased coronary artery. Inter-plaque stresses and plaque-adjacent intimal stresses are known to be correlated with location-specific rupture risk ⁴⁷.

The GIBBON toolbox^48,49 was used to define the material values, boundary conditions, and pressure surfaces for FE simulations as well as input formatting for FEBio⁵⁰ (Fig. 9). Each of the tissue types in the output plaques were 1-term⁵¹, 2-term^52,53 or 3-term¹⁷ Ogden solids⁵⁴. The general Ogden model has a volumetric component $U\left(J\right)$ where $J$ is the Jacobian modulus and a deviatoric component with $N=2$ terms, principle stretch ratios ${\lambda }_i$ with material moduli $c_i$ and exponent constants $m_i$^53,55.

$$\begin{array}{c}W\left({\lambda }_1,{\lambda }_2,{\lambda }_3,J\right)=\sum _{i=1}^N\frac{c_i}{{m_i}^2}\left({{\tilde{\lambda }}_1}^{m_i}+{{\tilde{\lambda }}_2}^{m_i}+{{\tilde{\lambda }}_3}^{m_i}-3\right)+U\left(J\right).\end{array}$$ #(1)

The plaques were evaluated to assess tissue stability under normal internal pressure ($P\approx 14.6$ kPa)⁵⁶, at pre-balloon angioplasty pressure ($P\approx 100$ kPa), and at the approximate upper limit of balloon angioplasty pressure ($P\approx 2.0$ MPa)⁵⁷. The pressure was loaded linearly over a minimum of 100 time steps (Fig. 10). The dynamic stress and strain relationships unique to patient-specific distributions of different tissues are essential for identifying how these distributions contribute to plaque stability⁵⁸. Allowance was made for a maximum of 25 stiffness matrix reformations with the optimal number of full-Newton solver iterations set to 15 per time step. Time steps that failed to converge were allowed a maximum of 3 retries to converge from the previous successful time step. Each subsequent retry was restarted at the last successful time-step and attempted again with a smaller time-step increment. This allowed for better discrimination between plaque models that were numerically unstable (e.g., due to meshing) for larger time-step increments and physiologically unstable plaques (Fig. 11).

Figure 9.

Finite element (FE) model of a synthetic coronary artery plaque with boundary conditions (black dots) and pressure surface (red transparent surface) (shown left) and material composition (shown right).

Figure 10.

Physiological robustness evaluation pipeline shown for 4 unique synthetic plaques. These plaques were all evaluated to be stable up to balloon angioplasty pressures (100 kPa) without rupture. Notice also that many of the high-stress regions are located along the intima (i.e., the pressure surface) . Lipid pools are seen to cause identifiable stress gradients in adjacent areas due to shearing while calcifications maintain their shape and do not.

Figure 11.

A sample cohort of the plaques from Fig. 7 as 3D models, some of which are not stable (e.g., extremely small lumen, dangerous stress gradient, etc.) as indicated by brighter intensities in the stress images.

At any of the simulated pressures, exceeding the stress failure threshold for any of the tissues (typically intima–lipid interface) resulted in the classification of the plaque as unstable. For the unstable plaques, the failure pressure was noted. The failure stress magnitude was $E\ge 300$ kPa, and constituted definite tissue failure and plaque rupture.⁵ Internal stress magnitudes $135\text{kPa}\le E<300\text{kPa}$ represented what Costopolous et al.⁵⁹ determined to be a sensitive and specific threshold in discriminating between probable ruptures and stable plaques. Plaques with internal stresses $E<135$ kPa under pressure loading were determined to be stable. Tissue type distributions as classified by CTA, such as simulated here, are also related to histological studies of the vulnerable plaque⁶⁰.

Pilot study combining plaque models with XCAT and DukeSim

The methods in the previous sections were used to generate and classify a library of 200 anatomically variable plaque models. To demonstrate the utilization of the plaque models to investigate emerging CT imaging techniques such as photon-counting CT, example plaques were inserted into an adult male XCAT³⁷ phantom, 50^th percentile in BMI. One plaque, containing a lipid-rich necrotic core with a single large calcification, was set to be 10 mm in length and was placed within the right coronary artery (RCA). A second plaque, containing a lipid-rich necrotic core with a smaller calcification, was set to be 18 mm in length and was placed within the left circumflex (LCX) artery. Each plaque model consisted of a segment of the coronary vessel plus the interior plaque. The different lengths, indicated above, were set according to typical measurements for coronary plaques⁶¹. Using the Rhinoceros modeling program, www.rhino3d.com, each plaque-containing vessel segment was manually placed along the desired XCAT coronary vessel in the desired location, Fig. 12. Contours were taken through the outer wall and lumen of the XCAT coronary vessel leading up to and after the plaque segment using the Section command within Rhinoceros. Contours were also taken through the corresponding vessel surfaces within the plaque segment. The contours were then lofted into new surfaces for the lumen and outer wall of the XCAT coronary vessel using the Loft command. Surfaces defining the plaque interior, calcifications and lipids, were smoothed using the Smooth command in Rhinoceros to taper their ends. The new surfaces for the coronary vessels (used to replace the original XCAT vessels) and the surfaces defining the plaques were input into the XCAT program as user-defined objects. The XCAT program allows users to add additional objects to the phantom that will move with any simulated motion.

Figure 12.

Incorporation of a plaque model within the right coronary artery of the XCAT phantom using the Rhinoceros modeling program. (Left) The surfaces of the plaque segment are manually placed along the vessel at the desired location. (Middle) Contours (yellow) are taken along the lumen and outer vessel surfaces of the original XCAT vessel as well as from the plaque vessel segment. (Right) The contours are lofted into new surfaces to replace those of the XCAT. Surfaces defining the plaque interior, such as the calcification in the above example, can be smoothed using the Smooth command in Rhinoceros to taper their ends.

The XCAT program was then used to simulate cardiac motion from the phantom at a heart rate of 60 BPM²⁷ (no respiratory motion was simulated). A series of 100 3D phantoms were output over time, centered over the mid-diastole phase, giving each 3D phantom a temporal resolution of 2.8 msec per phantom instance. The phantoms were output as voxelized images with an isotropic resolution of 0.1 mm.

The time series of phantoms was then input to the DukeSim³¹ simulator to undergo CT simulation emulating a single-source energy-integrating CT system (Siemens Flash) and a single-source photon-counting CT system (Siemens NEAOTOM Alpha) at 120 kV and 350 mAs with a pitch of 1 and a gantry rotation speed of 280 ms. The mid-diastole motion modeled by the phantoms was used in the simulation to approximate an ECG-gated scan. The 3D instances of the phantom were temporally synchronized with the CT simulator, selecting a phantom from the 100 frames to match time wise with each projection (1152 projections for energy-integrating CT and 2016 projections for photon-counting CT). The simulated CT imaging was performed with and without an iodinated contrast material at the equivalent of 370 mgI/mL (with contrast-enhanced CT values of approximately 250 HU in coronary arteries). Images were reconstructed with vendor-specific software (ReconCT, Siemens) using the Br40 reconstruction kernel. Slice thicknesses in reconstructed images were 0.2 mm for the photon-counting CT images and 0.75 mm for the energy-integrating CT images. This phantom simulation focused on a qualitative comparison of plaque appearance between photon counting detector CT (PCD) and conventional energy-integrating detectors (EID). The simulated images were directly compared for each of the three lesions on equivalent slices of each reconstructed image.

III. Results

Software

The complete software required to produce and assess the synthetic plaques was developed as standalone Podman⁶² and Docker⁶³ images with the minimum necessary runtime environment (totaling $\sim 10$ GB as a compressed archive). The DC-GAN software was run with one discrete card from an NVIDIA Tesla K80 GPU board, taking roughly 1 hour to train the model as described in Section 2.2. Subsequent use of the trained model to produce 200 new plaque images took $<2$ seconds with the GPU but still only required $\sim 1$ minute to do so with a CPU. Using CPU, the segmentation process took $<5$ minutes for 200 plaques. FE simulation required up to 2 hours per plaque for the meshing parameters (containing an average of $\sim {10}^5$ tetrahedra) used in Section 2.3.

To increase the interoperability of the synthetic plaque generation tools (see Sections 2.2–2.3) with separate software packages, the full-time sequence of FE simulation results were saved in the .feb and .xplt file formats (for ease of use at any of the simulated pressures up to failure with FEBioStudio). Additionally, the deformed node set was used to produce a complete set of .node .face, .ele, and .edge files (for ease-of-use with Tetgen, MEDIT, VTK, etc.), as well as plain-text files containing the locations of the nodes of the tetrahedra (a 4-column double-precision array) and the order in which those nodes connect to adjacent nodes to form faces (a 4-column double-precision array without material encoding or a 4-column double-precision array + 1 unsigned 8-bit integer column vector to encode material).

For accessibility, the software tool was written such that each of the steps of the process can be run individually manually from inside the running container if desired but can also be automated entirely using an example bash script and example configuration file included with the Docker image. Additionally, output files were sorted into top-level directories for each process stage. Upon completing the process, these directories were also automatically subdivided by data type (e.g., .ply, .png, etc.). The directories containing any data unique to the user session were automated to copy from the docker container to the directory on the host operating system from which the Docker container was started.

3D plaque models

The 3D plaque models were output in several file formats representing regions of different vessel tissue types. Solid (i.e., tetrahedral) meshes—as used in finite element analysis in Fig. 10–11 above—and as triangulated surfaces (i.e., one 2-manifold, watertight .STL file for each tissue type present in the vessel and plaque)—as used in computational phantom simulation work and shown in Fig. 13.

Figure 13.

Example plaques represented by triangulated surfaces saved as .STL files. Multiple plaques are shown in (A) with a single plaque highlighted in (B). Unremarkable materials included are the adventitia, lumen, and intima. Abnormalities included are lipid-rich necrotic core (LRNC), intraplaque hemorrhage (IPH), calcification (CALC), and fibrous matrix (MATX).

PCD-CT and EID-CT simulation

Photon counting CT simulations and the equivalent EID CT simulations are shown in Fig. 14. Results are shown for the plaque placed in the RCA with and without contrast for both simulations; similar results were found for the plaque located within the LCX (results not shown). Overall, it can be observed that PCD-CT has potential to better resolve the plaque and contains more high-frequency content than the EID-CT images. The simulations provide an initial demonstration of how the plaque models generated in this work can be used as targets to compare imaging devices and techniques, enabling future detailed comparisons.

Figure 14:

(Left) Transaxial image slice of the XCAT heart with the plaque inserted in the right coronary artery. The calcium and lipid components of the plaque are labeled along with the lumen of the RCA. (Middle and right) Non-contrast and contrast enhanced CT simulations with energy integrating detectors (EID) and photon counting detectors (PCD) showing the corresponding imaging-based renditions of the plaque zoomed in on the region indicated by the green box in the phantom image. The calcium component of the plaque appears white in the images. The lipid component appears more visible in the PCD-CT images, appearing as a dark spot next to the calcium as indicated by the white arrows.

IV. Discussion

In this work, we developed a method to synthesize a library of realistic coronary plaques based on micron-level resolution plaque histology and morphology data. The plaques, containing different tissue types, can be combined with the library of XCAT phantoms to simulate any number of cases for use in virtual imaging trials in cardiac CT imaging. We have also presented initial simulations of these atherosclerotic plaques combined with XCAT phantoms with our EID-CT³¹ and PCD-CT⁶⁴ simulator to illustrate their utility.

We unified tools from prior work^{26,27,30,36,37,64–66} to create a framework for reproducible, simulated cardiac imaging of realistic virtual coronary artery pathologies. These tools enable cardiac CT optimization by creating a framework for easy parameterization of realistic cardiac motion in XCAT, integrated CAD models, and CT acquisition and reconstruction—essential for optimizing and determining the robustness of image acquisition and reconstruction parameters. The results demonstrate the fundamental limitations of acquiring, quantifying, and determining the composition of CAD from cardiac CT with unoptimized acquisition protocols. With realistic coronary artery plaques, physiologically realistic cardiac motion and anatomy, and a physics-based CT simulator, this work enables future optimization of image acquisition and reconstruction parameters as a function of plaque composition, plaque location, percent-stenosis, heart rate, patient body habitus, the likelihood of plaque rupture (as indicated by plaque composition and location), among others.

This work provides an EID-CT and PCD-CT simulation-compatible tool for generating and validating cohorts of coronary artery plaques with realistic morphology and tissue composition having up to 4 types of abnormal tissue per plaque. While prior work has produced models that consider progressive tissue injury over time from tissue loading cycles, or detailed analysis of anatomically at-risk locations, there were no previous coronary artery plaque models that could be used with an existing computational phantom cohort (XCAT) and an existing CT simulation platform (DukeSim). The combination of XCAT phantom integration and DukeSim integration is intended to streamline interoperability and enable its use in computational imaging and clinical parameter optimization studies. Direct comparisons between and across different heartrates, plaques (locations and orientations), image acquisition parameters, and reconstruction parameters can be made. Additionally, this tool enables work that can directly quantify distinctions between image quality in simulated-based EID-CT and photon-counting CT.

This work has several limitations. First, users of the plaque generation tool cannot currently control the characteristics of the generated plaques. To get a subset of plaques with desired characteristics, a user would need to generate many plaque images and sort through them. Future work will investigate additional processing to help guide the output of the tool; for example, targeting different percentages of the various plaque components. We will also investigate presorting the histology images into different classifications to train separate networks. Second, the plaque generation tool currently outputs only a single 2D image of a plaque. Extrusion is then used to extend the 2D image into a 3D model. We are currently working to get additional histological data of plaques, containing multiple slices taken along their length, and using these to train a 3D deep learning network. Third, plaques embedded into the XCAT coronary arteries are not validated to have a realistic orientation with respect to the long axis of the vessel; this is due to the unspecified orientation of the arterial cross-sections in the original training data and further compounded by the data augmentation process in which arbitrary rotations of plaques were included. Additionally, the CT simulation work was limited to just two example plaques over one phase of the heart under a single set of imaging parameters for PCD-CT and EID-CT. Future work will involve a more extensive study. Finally, the histology images used for training were from excised carotid arteries rather than coronary arteries. As mentioned above, we are investigating obtaining additional histological data; this includes data obtained from the coronaries.

V. Conclusion

Plaques generated using our techniques demonstrate good anatomical agreement and robust physiological agreement with clinical plaques. The framework we developed generates realistic and variable CAD pathologies. These pathologies can serve as a known truth to evaluate and optimize cardiovascular imaging devices and applications through full virtual imaging trials.

Data Sharing

Authors are not able to share data at this time.