Advertisement

Atherosclerosis risk classification with computed tomography angiography: A radiologic-pathologic validation study

Open AccessPublished:November 23, 2022DOI:https://doi.org/10.1016/j.atherosclerosis.2022.11.013

      Highlights

      • Machine learning software based on histologically defined tissue characterization inputs demonstrated high accuracy for plaque risk phenotype determination compared to matched 2 mm tissue sample cross-sections: weighted kappa 0.82 and 0.70 in derivation and validation cohorts.
      • The area under the receiver operating curve for software correct identification of plaque type was 0.97, 0.95, 0.99 for unstable plaque, stable plaque, and minimal disease, respectively.
      • Percent lumen diameter stenosis per cross-section correlated poorly with plaque risk phenotypes.

      Abstract

      Background and aims

      The application of machine learning to assess plaque risk phenotypes on cardiovascular CT angiography (CTA) is an area of active investigation. Studies using accepted histologic definitions of plaque risk as ground truth for machine learning models are uncommon. The aim was to evaluate the accuracy of a machine-learning software for determining plaque risk phenotype as compared to expert pathologists (histologic ground truth).

      Methods

      Sections of atherosclerotic plaques paired with CTA were prospectively collected from patients undergoing carotid endarterectomy at two centers. Specimens were annotated for lipid-rich necrotic core, calcification, matrix, and intraplaque hemorrhage at 2 mm spacing and classified as minimal disease, stable plaque, or unstable plaque according to a modified American Heart Association histological definition. Phenotype is determined in two steps: plaque morphology is delineated according to histological tissue definitions, followed by a machine learning classifier. The performance in derivation and validation cohorts for plaque risk categorization and stenosis was compared to histologic ground truth at each matched cross-section.

      Results

      A total of 496 and 408 vessel cross-sections in the derivation and validation cohorts (from 30 and 23 patients, respectively). The software demonstrated excellent agreement in the validation cohort with histological ground truth plaque risk phenotypes with weighted kappa of 0.82 [0.78–0.86] and area under the receiver operating curve for correct identification of plaque type was 0.97 [0.96, 0.98], 0.95 [0.94, 0.97], 0.99 [0.99, 1.0] for unstable plaque, stable plaque, and minimal disease, respectively. Diameter stenosis correlated poorly to histologically defined plaque type; weighted kappa 0.25 in the validation cohort.

      Conclusions

      A machine-learning software trained on histological ground-truth tissue inputs demonstrated high accuracy for identifying plaque stability phenotypes as compared to expert pathologists.

      Graphical abstract

      Keywords

      1. Introduction

      Cardiovascular disease is the most common cause of death and disability globally, driven mainly by myocardial infarction and ischemic stroke from unstable atherosclerosis [
      World Health Organization (WHO)
      Cardiovascular diseases (CVDs) fact sheet. 2017 23.
      ]. It exerts an exorbitantly high financial burden on society [
      • Bloom D.E.
      • Cafiero E.T.
      • Jané-Llopis E.
      • Abrahams-Gessel S.
      • Bloom L.R.
      • Fathima S.
      • Feigl A.B.
      • Gaziano T.
      • Mowafi M.
      • Pandya A.
      • Prettner K.
      • Rosenberg L.
      • Seligman B.
      • Stein A.Z.
      • Weinstein C.
      ]. Patient risk management largely depends on population-based scoring methods or secondary prevention in patients with established disease [
      • Lyngbakken M.N.
      • et al.
      Novel biomarkers of cardiovascular disease: applications in clinical practice.
      ,
      • Hafiane A.
      Vulnerable plaque, characteristics, detection, and potential therapies.
      ]. Developing diagnostics for more precise patient risk categorization and tailored therapeutics is an area of significant clinical investigation. Despite discoveries of novel predictive plasma biomarkers [
      • Lyngbakken M.N.
      • et al.
      Novel biomarkers of cardiovascular disease: applications in clinical practice.
      ], routine diagnostic methods for identifying individuals and lesions at high risk for atherothrombosis in coronary or extracranial arteries are still lacking [
      • Hafiane A.
      Vulnerable plaque, characteristics, detection, and potential therapies.
      ]. Despite considerable scholarship establishing that the high-risk (unstable) atheroma is characterized by several structural and biological features [
      • Buckler A.J.
      • et al.
      Virtual transcriptomics: noninvasive phenotyping of atherosclerosis by decoding plaque biology from computed tomography angiography imaging.
      ] with greater dependence on plaque tissues than luminal narrowing [
      • Falk E.
      • et al.
      The high-risk plaque initiative: primary prevention of atherothrombotic events in the asymptomatic population.
      ], clinical guidelines generally base care on the measurement of arterial lumen diameter stenosis due to lack of an objectively defined and validated alternatives that incorporate histologically-validated plaque characteristics.
      Whereas a number of radiological modalities have strong merits for cardiovascular imaging, CTA is an established, frequently utilized test with the benefit that it is applicable to multiple arterial beds [
      • Abdelrahman K.M.
      • et al.
      Coronary computed tomography angiography from clinical uses to emerging technologies: JACC state-of-the-art review.
      ,
      • Saba L.
      • et al.
      Roadmap consensus on carotid artery plaque imaging and impact on therapy strategies and guidelines: an international, multispecialty, expert review and position statement.
      ]. However, the evaluation of plaque risk (risk of rupture or erosion) relies on descriptive assessment based on Hounsfield unit (HU) appearance and arterial remodeling. Given that luminal stenosis remains the dominant measure of disease severity used in clinical practice, we also sought to assess the agreement between diameter stenosis and histological plaque stability. Specifically, here we use stenosis as a comparator to indicate the effectiveness of stenosis as well as our method in the same patients in terms of their respective abilities to determine plaque stability. Recently, machine-learning approaches have demonstrated improvements in event prediction from CTA images based on plaque quantification [
      • Munger E.
      • et al.
      Application of machine learning in understanding atherosclerosis: emerging insights.
      ,
      • Zhang L.
      • et al.
      Predicting locations of high-risk plaques in coronary arteries in patients receiving statin therapy.
      ,
      • Han D.
      • et al.
      Machine learning framework to identify individuals at risk of rapid progression of coronary atherosclerosis: from the PARADIGM registry.
      ]. Still, the relationship of this approach to established histologic plaque phenotypes, such as thin-capped fibroatheromas, has not been performed. Techniques such as [
      • Hell M.M.
      • et al.
      Quantitative global plaque characteristics from coronary computed tomography angiography for the prediction of future cardiac mortality during long-term follow-up.
      ,
      • Han D.
      • et al.
      Prognostic significance of plaque location in non-obstructive coronary artery disease: from the CONFIRM registry.
      ] have hitherto offered an assessment by arbitrary HU cutoffs unsubstantiated by histology standards, are influenced by imaging protocol parameters, and do not allow discrimination of overlapping density tissues such as lipid-rich necrotic core (LRNC) and intra-plaque hemorrhage (IPH). These prior methods have not utilized histological ground truth and, whereas they seek to measure aspects of plaque morphology, stop short of classifying the phenotype. Recently, the RSNA's Quantitative Imaging Biomarker Alliance (QIBA) recommended that imaging-based tissue characterization should ideally be based on reference truth according to histology as the ground truth [].
      We sought to validate a machine-learning approach to identify high-risk plaque across an entire vessel according to an objective definition accepted widely in cardiovascular medicine and validated in histology, which we term histology-defined high-risk plaque (HD-HRP) for clarity. We also compared its performance relative to the ability of lumen stenosis to determine plaque stability.

      2. Materials and methods

      2.1 Study design and patient data source

      This prospective study (NCT02143102) enrolled consecutive adults from two centers (Louisiana State University Health Sciences Center, New Orleans [urban center], LA, and West Jefferson Medical Center, Marrero, LA [suburban center]) who were scheduled to undergo cardiovascular CT angiography (CTA) and subsequent surgical carotid endarterectomy within 30-days of enrollment, as previously described [
      • Sheahan M.
      • et al.
      Atherosclerotic plaque tissue: noninvasive quantitative assessment of characteristics with software-aided measurements from conventional CT angiography.
      ]. The two centers represent different patient demographics, allowing the opportunity to test generalizability. Each patient provided informed, written consent, and the local institutional review boards approved the study protocol at each participating center. Model evaluation utilized a separate derivation and validation set assigned using sequential enrollment, which contained unseen data.

      2.2 Histopathologic examination

      Surgical histological specimens were processed at the Louisiana State University Health Sciences Center [
      • Sheahan M.
      • et al.
      Atherosclerotic plaque tissue: noninvasive quantitative assessment of characteristics with software-aided measurements from conventional CT angiography.
      ]. Specimen preparation and sectioning were performed at the Louisiana State University (LSU) School of Medicine's Pathology department, with cross-sectional specimens prepared at every 2.0 mm throughout the length of the endarterectomy specimen. Histopathological assessment was performed at CV Path Institute (Gaithersburg, MD), resulting in annotated histology images using the ZEN software tool (Zeiss, Oberkochen, Germany), independently by three pathologists (RV, LG, AS) who were blinded to clinical and CT data. The validation cohort was annotated by all three, and combinations of 2 annotators each annotated the derivation cohort. The final interpretation was based on consensus (2 of 3 agreement). The atherosclerotic plaque was characterized independently by each reader at each 2.0 mm section. The American Heart Association have approved this phenotype system to help inform practicing clinicians in determining what course of treatment may benefit their patients, based on extensive scholarship that has progressively been refined over the last 20 years. This typing system was originated by Stary [
      • Stary H.C.
      Natural history and histological classification of atherosclerotic lesions: an update.
      ,
      • Stary H.C.
      • et al.
      A definition of advanced types of atherosclerotic lesions and a histological classification of atherosclerosis A report from the Committee on Vascular Lesions of the Council on Arteriosclerosis, American Heart Association.
      ] and refined by Virmani by making a closer tie to events [
      • Virmani R.
      • et al.
      Pathology of the vulnerable plaque.
      ,
      • Virmani R.
      • et al.
      Lessons from sudden coronary death: a comprehensive morphological classification scheme for atherosclerotic lesions.
      ]. This system and other terms used in various reports are summarized in Table 1, indicating the banding into the three phenotypes detected by this work, minimal (which includes normal), stable, or unstable, with histopathology-verified classification agreement (Fig. 1, Fig. 2, Fig. 3). Broad clinical consensus on these definitions will benefit from experience, with updates in the future as warranted.
      Table 1Histological plaque stability typing system.
      Virmani systemOther terms for the same lesions
      MinimalIntimal xanthomaFatty dot or streak, early lesion
      Intimal thickening
      StablePathological intimal thickening
      Calcified noduleCalcified plaque
      Fibro-calcific plaqueFibrous plaque
      UnstableUlcerationAtheromatous plaque
      IPHFibrolipid plaque
      TCFAComplicated lesion
      Fibroatheroma
      Rupture
      Fig. 1
      Fig. 1Minimal presentations of atherosclerosis initiation.
      Samples are selected from the clinical study Non-invasive Computer-Aided Phenotyping of Vasculopathy (Q-CAMP), NCT02143102, with Movat pentachrome stain. Fragments representing minimal phenotypes are generally pieces at the end of more clinically significant specimens. Neovascularization in xanthoma are color-coded in purple. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)
      Fig. 2
      Fig. 2Stable presentations of atherosclerosis initiation.
      Samples selected from the clinical study Non-invasive Computer-Aided Phenotyping of Vasculopathy (Q-CAMP), NCT02143102, with Movat pentachrome stain. Colors: LRNC, yellow; Calcification, aquamarine; Neovascularization, purple; Ulceration, green; IPH, red; Lipid pool, orange. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)
      Fig. 3
      Fig. 3Unstable presentations of atherosclerosis initiation.
      Samples selected from the clinical study Non-invasive Computer-Aided Phenotyping of Vasculopathy (Q-CAMP), NCT02143102, with Movat pentachrome stain. Colors: LRNC, yellow; Calcification, aquamarine; Neovascularization, purple; Ulceration, green; IPH, red; Lipid pool, orange. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

      2.3 CT image acquisition and CT-based tissue characterization

      CTA acquisitions were performed according to local imaging protocols using a variety of CT scanner types (16–256 row), manufacturers, and acquisition parameters. Imaging was performed with 0.75–2.5 mm sections, 120 kVp, 267–600 mAs, and six different reconstruction kernels according to vendor and site-specific protocols. CTA datasets were analyzed using commercially available quantitative atherosclerotic plaque analysis software (ElucidVivo, Elucid Bioimaging, Boston, MA) [
      • Buckler A.J.
      • et al.
      Virtual transcriptomics: noninvasive phenotyping of atherosclerosis by decoding plaque biology from computed tomography angiography imaging.
      ,
      • Abdelrahman K.M.
      • et al.
      Coronary computed tomography angiography from clinical uses to emerging technologies: JACC state-of-the-art review.
      ,
      • Sheahan M.
      • et al.
      Atherosclerotic plaque tissue: noninvasive quantitative assessment of characteristics with software-aided measurements from conventional CT angiography.
      ,
      • Varga-Szemes A.
      • et al.
      Coronary plaque assessment of Vasodilative capacity by CT angiography effectively estimates fractional flow reserve.
      ,
      • Zhu G.
      • et al.
      Semiautomated characterization of carotid artery plaque features from computed tomography angiography to predict atherosclerotic cardiovascular disease risk score.
      ,
      • Chrencik M.T.
      • et al.
      Quantitative assessment of carotid plaque morphology (geometry and tissue composition) using computed tomography angiography.
      ,
      • Rafailidis V.
      • et al.
      Carotid Plaque Vulnerability: the Correlation of Plaque Components as Quantified Based on Computed Tomography Angiography with Neurologic Symptoms.
      ,
      • Gupta A.
      • et al.
      Semi-automated detection of high-risk atherosclerotic carotid artery plaque features from computed tomography angiography.
      ,
      • van Assen M.
      • et al.
      Automated plaque analysis for the prognostication of major adverse cardiac events.
      ]. Custom software (HistoMatch, Elucid Bioimaging) matched histologic sections with corresponding locations on the CT image.
      The software processed CTA images with an image restoration approach based on measuring the patient-specific 3-dimensional point spread function (PSF) local to the vessel walls. This PSF was used within the tissue characterization as a spatially varying deconvolution to enhance the effective spatial and contrast resolution such that image intensities are restored to more closely represent the original materials imaged. This not only mitigates visually conspicuous calcium blooming even when less visually apparent. The vessel wall was analyzed, defining plaque into different components: lipid-rich necrotic core (LRNC), calcification (CALC), intraplaque hemorrhage (IPH), matrix (MATX), perivascular adipose tissue (PVAT), fibrous cap thickness (the smallest distance from LRNC to the lumen). See Supplementary Methods for details on the tissue characterization. Degree of stenosis was computed based on the ElucidVivo lumen segmentations.

      2.4 Plaque stability phenotype classification

      Plaque phenotyping was, as reported here, an extension to the prior version of the software. It was performed at each artery cross-section using 2 mm intervals to match the histopathologic approach. The model architecture is a Resnet-18 with partially frozen layers, a CNN block, and a fully connected layer. Model inputs are formed by presenting structural anatomic and tissue characteristics from the image processing in the form of enriched images. The enriched images are formatted within a normalized coordinate system to exclude size or arterial bed differences not pertinent to stability while enhancing those aspects of the morphology which does. The pre-trained partially frozen Resnet-18 and the CNN block act as feature detectors, and the fully connected layers act as the classifier. The initial layers of pre-trained Resnet-18 are frozen while the rest are unfrozen to fine-tune the model to learn the task of phenotype classification. The number of frozen layers is set as a hyperparameter. The output of the HD-HRP deep learning model for a particular cross-section was three likelihoods of the cross-section being minimal, stable or unstable plaque phenotype, with the largest number among the three likelihoods chosen as the predicted phenotype (Supplementary Figs. S1 and S2). We use likelihoods rather than probabilities as an appropriate way to recognize that the three numbers are not constrained to sum to 1, as would be the case with probabilities. The biophysiological interpretation is that each is determined without constraint of the others to account for the complexity of the varied presentations of human disease. The inference is processed in under 5 min for a full arterial tree using the quantitative morphology inputs.

      2.5 Statistical analysis

      Clinical performance testing of anatomic structure calculations proceeded in two steps. First, the quantitative assessment of structural anatomy and tissue characterizations were compared against histology. Second, HD-HRP phenotype classification (minimal, stable, unstable) was assessed versus histology. The accuracy of model input measurements (plaque tissues) is a characterization of technical performance. The agreement with expert pathologists vs. the model output for classifying the phenotype is considered a clinical performance evaluation.
      Correctness of classification was assessed relative to expert labeling performed blinded to the imaging, and the imaging model performed blinded to the reference truth. These metrics penalize model performance with increasing weight for misclassifications that occur further from the true classification. The test set results are summarized in a three-by-three confusion matrix for validation. We considered the three-by-three confusion table of their phenotype rating results to assess agreement among the pathologists, calculating the corresponding weighted kappa statistic using the quadratic weights. The relationship between diameter stenosis (as a continuous-valued measurement) for predicting plaque stability phenotypes was also explored. The weighted kappa statistic with a 95% confidence interval was used to quantify the overall strength of agreement between the multi-category classification of the predictive modeling compared with the histopathology reference result. A quadratic weighting scheme described the closeness of agreement between the categories. Generalizability was established using two independent sets of the validation data, one conducted in an urban center and independently on data from a suburban center, to demonstrate performance in disparate subpopulations.

      3. Results

      The study group comprised 53 subjects with a mean age of 59.7 years (57% male, n = 18 for urban and 35 for suburban). Thirty subjects were assigned to the derivation cohort and 23 to the validation cohort based on sequential enrollment. Baseline characteristics of the study population are shown in Supplementary Table S1. We noted that the two centers represented different relevant demographic and risk factor profiles. The urban center was significantly less white and with lower prevalence of hypercholesterolemia. Agreement between pathologists is summarized in Supplementary Table S2. We noted that the pathologists generally agreed at the banded level, but one pathologist tended to differ on the granular phenotype. Overall, there was good agreement between readers, with the final section interpretation based on consensus (2 of 3). There was a total of 496 and 408 vessel cross-sections in the development and validation cohorts, respectively. According to histological ground truth, there were 168, 160, and 168 minimal, stable, and unstable sections in the development cohort and 159, 94, and 155 minimal disease, stable plaque, and unstable plaque sections in the validation cohort.
      The performance of the study software in the validation cohort compared to pathologist consensus (histologic ground truth) is shown in Table 2 and Fig. 4, stratified by overall study cohort and center (urban and suburban). The HD-HRP deep learning model was found to perform at a high level across developmental and validation cohorts and among subjects from each participating center, with weighted kappa values showing excellent agreement. There were no instances where the software predicted an unstable phenotype, and histology showed minimal disease. In 13 instances, the software predicted unstable phenotype, and histology showed stable disease. The area under the receiver operating curve for correct identification of plaque type was 0.97 [0.96, 0.98], 0.95 [0.94, 0.97], 0.99 [0.99, 1.0] for unstable plaque, stable plaque, and minimal disease, respectively.
      Table 2Study software performance of plaque phenotyping versus histologic ground truth and confusion matrices stratified by center.
      Development PerformanceValidation Performance
      CombinedUrban CenterSuburban CenterCombinedUrban CenterSuburban Center
      Agreement with Expert Pathologist0.93 [0.91,0.95]0.93 [0.91,0.95]0.93 [0.91,0.95]0.9 [0.87,0.93]0.92 [0.87,0.96]0.89 [0.85,0.93]
      Kappa0.89 [0.86,0.92]0.89 [0.86,0.92]0.89 [0.86,0.92]0.85 [0.81,0.89]0.86 [0.78,0.94]0.83 [0.78,0.89]
      Weighted Kappa0.82 [0.78,0.86]0.82 [0.78,0.86]0.82 [0.78,0.86]0.7 [0.62,0.79]0.69 [0.57,0.81]0.66 [0.55,0.77]
      Confusion Matrix
      CombinedUrban CenterSuburban Center
      TruePredictedCountTruePredictedCountTruePredictedCount
      stablestable75stablestable4stablestable68
      minimalstable0minimalstable0minimalstable0
      unstablestable21unstablestable4unstablestable12
      stableminimal6stableminimal2stableminimal1
      minimalminimal159minimalminimal45minimalminimal114
      unstableminimal0unstableminimal0unstableminimal0
      stableunstable13stableunstable5stableunstable14
      minimalunstable0minimalunstable0minimalunstable0
      unstableunstable134unstableunstable72unstableunstable67
      Fig. 4
      Fig. 4(Left panel): The results are quantitatively and pictorially applied to a left coronary artery system. The 3D image represents the results along the centerline, and the rectangular plots are the likelihoods produced by the CNN along the vessel length. In these plots, green signifies “minimal” disease, yellow represents “stable” plaque, and red indicates “unstable” plaque. The y-axis of the graphs is the likelihood of these three phenotypes. (Right panel): Receiver Operating Characteristic Curves for the three types. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)
      The severity of diameter % stenosis on CTA on a section-basis showed poor agreement with plaque risk classification according to pathologic consensus. Specifically, diameter stenosis agreed with expert pathologist assessment of plaque risk in only 49% of sections, with weighted kappa of 0.27 [0.24–0.3] and 0.25 [0.21–0.28] in the development and validation cohorts, respectively.

      4. Discussion

      Plaque disruption (rupture or erosion) is the most common cause of arterial thrombotic events such as stroke and myocardial infarction. Cardiovascular CTA is an increasingly utilized test for assessing patients with known or suspected cardiovascular diseases. While assessing stenosis on CTA serves as a measure of disease severity and risk and guides management in many patients following current guideline-based care, there is significant interest in further improving CTA's diagnostic and prognostic yield by incorporating measures of plaque stability. Current qualitative approaches using measurements of remodeling, calcification patterns, and Hounsfield unit thresholds to assess plaque risk and volume are currently limited by the absence of comparison to histological ground truth. In the current prospective study, we overcome these limitations by evaluating the performance of a commercially available, FDA-cleared, interpretable machine-learning software for identifying plaque risk on CTA versus histological ground truth.
      The findings from the current study demonstrated that there is excellent agreement with plaque classification on CTA as compared to expert pathologists. Further, this generalizes across two different demographics as represented by different care settings. This study extends prior work demonstrating that the commercially available software used in this study was highly accurate compared to histologic ground truth for measuring plaque tissue characteristics and volumes, including quantifying calcified, LRNC, and matrix plaque components (12). Compared to prior studies using this software, the current study includes a larger sample size and extends the comparison to histology to include plaque risk phenotype classifications. It is important to demonstrate the level of agreement using histology as ground truth. Future studies assessing the prognostic value of CT-based tissue (plaque) typing will seek to evaluate this imaging approach's prognostic and clinical implications in larger, diverse patient populations. Further, studies assessing serial changes in atherosclerotic tissues related to pharmacotherapies or other interventions will benefit from establishing the accuracy of CT-based plaque phenotyping, as demonstrated in this study.
      In addition, this study demonstrated an overall poor relationship between lumen stenosis and plaque stability when assessed on a per cross-sectional basis. This finding highlights the additive and complementary role that HD-HRP determination may play in refining patient risk and potentially better-informing treatment strategies. For example, measures of LRNC or IPH despite “optimal” medical therapy using traditional agents (e.g., aspirin, statins) may be markers of disease progression and prompt the application of more advanced lipid or inflammatory modulating treatments. Additionally, information on HD-HRP may allow physicians and patients to better understand the biology of their atherosclerotic process and potential residual clinical risk.
      Given that we used carotid artery specimens harvested in endarterectomy procedures as a representative artery, it may be questioned whether image recognition validated against carotid tissue applies to plaque in other vascular beds, particularly the coronary. This is important because tissue collected from live patients with contemporaneous CTA is superior to autopsy specimens with a longer time difference in when the CTA is performed, and the tissue changes post-mortem that would otherwise result in lower quality comparisons. As such, the carotid artery would be considered not only an acceptable but even an ideal model artery, given that the approach used here has been engineered to exploit the commonalities while mitigating the difference across beds.
      First, technical factors such as how superficial the arteries are and motion are accounted for in accepted imaging protocols, having been optimized for human interpretation and which our algorithm uses to its advantage. Second, while the prevalence of tissues differs across arterial beds, the density distributions indicative of tissue type as used by the algorithms share common definitions. Plaque characteristics such as a large atheromatous lipid-rich core, thin fibrous cap, outward remodeling, infiltration of the plaque with macrophages and lymphocytes, and thinning of the media predispose to thrombosis in both carotid and coronary artery disease [
      • Sigala F.
      • et al.
      Coronary versus carotid artery plaques. Similarities and differences regarding biomarkers morphology and prognosis.
      ]. Third, tissues in the carotid and coronary arteries also have many similarities related to vascular tone regulation and pathological atherosclerotic response to changes in shear stress [
      • Sigala F.
      • et al.
      Coronary versus carotid artery plaques. Similarities and differences regarding biomarkers morphology and prognosis.
      ]. Clinical studies in both the coronary [
      • Chatzizisis Y.S.
      • et al.
      Association of global and local low endothelial shear stress with high-risk plaque using intracoronary 3D optical coherence tomography: introduction of 'shear stress score'.
      ] and the carotid arteries [
      • Gnasso A.
      • et al.
      In vivo association between low wall shear stress and plaque in subjects with asymmetrical carotid atherosclerosis.
      ] identify lower wall shear stress associated with plaque development and localization according to a common mechanism. These points support the conclusion that whereas the extent of the various plaque tissues differs across arterial beds, the cellular and molecular level milieu of the individual tissue types share common objective definitions across beds [

      Ibrahimi, P., et al., Coronary and carotid atherosclerosis: how useful is the imaging? Atherosclerosis. 231(2): p. 323-333.

      ]. Whereas inter-bed differences such as a thicker fibrous cap and a higher prevalence of intra-plaque hemorrhage in carotid vs. the coronaries affect the amount of these tissues [
      • Schaar J.A.
      • et al.
      Terminology for high-risk and vulnerable coronary artery plaques. Report of a meeting on the vulnerable plaque, June 17 and 18, 2003, Santorini, Greece.
      ], it does not change the nature of the tissues when they do present clinically or via diagnostic imaging. It is the cellular and molecular level milieu on which the algorithm operates, not the prevalence, enabling our choice of tissue model for tissue characterization.
      It is important to note the differences in the machine-learning approach utilized in this study compared to many papers in this area. We used a pipeline approach consisting of 2 stages with outputs that are individually objectively capable of validation at the biological level to feed the convolutional neural network. This approach aligns with the methodology for using radiologic images to serve as true biomarkers of risk as defined by the RSNA QIBA initiative. Specifically, this approach using CTA provides a classification scheme to determine phenotype non-invasively, where the classifier is trained on known ground truth defined outside of radiology rather than using more subjective assessments idiosyncratic to scanning protocol or interpreter. This also has the effect of increasing the leverage for each sample, increasing the signal-to-noise ratio in the data presented to the network in a manner that exceeds typical approaches. Coupled with this being the most extensive and detailed set of paired specimens with CTA having rich annotation at each of the two levels having yet been reported, sample size limitations have been effectively mitigated.

      4.1 Limitations and outlook

      Our study has several limitations. First, clinical outcomes of the procedure or subsequent clinical events were not assessed. Future study is needed to evaluate the incremental improvement in prognostic usefulness of CTA using this software compared to clinical variables, traditional CTA measures of plaque risk (stenosis, remodeling, attenuation), and plaque volume. Secondly, the classification into three risk groups may oversimplify the individual and plaque-specific risk and requires further study to assess its prognostic usefulness and the impact of medications and lifestyle changes on plaque evolution according to this classification scheme. We are optimistic that future studies using CTA measures of plaque volume and plaque risk phenotype as an endpoint in clinical trials may benefit from incorporating HD-HRP as an imaging biomarker if proven to predict clinically relevant outcomes independently but acknowledge that this will benefit by additional studies in each arterial bed.
      The current status establishes that our HD-HRP model could potentially be deployed in clinical CTA workflow. The classifier is embedded in the software with support for two deployment options supported, both analyses conducted via a SaaS model and support for on-site use directly by clinicians. Image preprocessing is self-contained within the software and accepts standard DICOM for CTA on all arterial beds.

      4.2 Conclusion

      In conclusion, systemic assessment of plaque stability is feasible and accurate from CTA when objectively validated tissue characterization is employed. Lumen percent diameter stenosis on CTA shows poor correlation with a histologically defined high-risk plaque on a per-segment analysis.

      Financial support

      Project funding was obtained in part by the  National Heart, Lung, and Blood Institute of the  National Institutes of Health  ( HL126224 ).

      CRediT authorship contribution statement

      Andrew J. Buckler: Conceptualization, Methodology, Writing – review & editing, Software. Antonio M. Gotto: Writing – review & editing, Software. Akshay Rajeev: Data curation, Software. Anna Nicolaou: Validation. Atsushi Sakamoto: Formal analysis. Samantha St Pierre: Data curation, Formal analysis. Matthew Phillips: Methodology, Software. Renu Virmani: Methodology, Writing – review & editing, Software. Todd C. Villines: Writing – review & editing.

      Declaration of competing interest

      The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:AJB and TCV are employees and shareholders of Elucid Bioimaging Inc. AR, AN, SS, and MP are employees of Elucid. AMG is an unpaid advisor to Elucid. AS and RV are employees of CVPath Institute.

      Appendix A. Supplementary data

      The following is the Supplementary data to this article:

      References

        • World Health Organization (WHO)
        Cardiovascular diseases (CVDs) fact sheet. 2017 23.
        (April 2020]; Available from:)
        • Bloom D.E.
        • Cafiero E.T.
        • Jané-Llopis E.
        • Abrahams-Gessel S.
        • Bloom L.R.
        • Fathima S.
        • Feigl A.B.
        • Gaziano T.
        • Mowafi M.
        • Pandya A.
        • Prettner K.
        • Rosenberg L.
        • Seligman B.
        • Stein A.Z.
        • Weinstein C.
        Forum W.E. The Global Economic Burden of Noncommunicable Diseases. 2011 (Geneva)
        • Lyngbakken M.N.
        • et al.
        Novel biomarkers of cardiovascular disease: applications in clinical practice.
        Crit. Rev. Clin. Lab Sci. 2019; 56: 33-60
        • Hafiane A.
        Vulnerable plaque, characteristics, detection, and potential therapies.
        J. Cardiovasc. Dev. Dis. 2019; 6
        • Buckler A.J.
        • et al.
        Virtual transcriptomics: noninvasive phenotyping of atherosclerosis by decoding plaque biology from computed tomography angiography imaging.
        Arterioscler. Thromb. Vasc. Biol. 2021; (Atvbaha121315969)
        • Falk E.
        • et al.
        The high-risk plaque initiative: primary prevention of atherothrombotic events in the asymptomatic population.
        Curr. Atherosclerosis Rep. 2011; 13: 359-366
        • Abdelrahman K.M.
        • et al.
        Coronary computed tomography angiography from clinical uses to emerging technologies: JACC state-of-the-art review.
        J. Am. Coll. Cardiol. 2020; 76: 1226-1243
        • Saba L.
        • et al.
        Roadmap consensus on carotid artery plaque imaging and impact on therapy strategies and guidelines: an international, multispecialty, expert review and position statement.
        AJNR (Am. J. Neuroradiol.). 2021; 42: 1566-1575
        • Munger E.
        • et al.
        Application of machine learning in understanding atherosclerosis: emerging insights.
        APL Bioeng. 2021; 5011505
        • Zhang L.
        • et al.
        Predicting locations of high-risk plaques in coronary arteries in patients receiving statin therapy.
        IEEE Trans. Med. Imag. 2018; 37: 151-161
        • Han D.
        • et al.
        Machine learning framework to identify individuals at risk of rapid progression of coronary atherosclerosis: from the PARADIGM registry.
        J. Am. Heart Assoc. 2020; 9: e013958
        • Hell M.M.
        • et al.
        Quantitative global plaque characteristics from coronary computed tomography angiography for the prediction of future cardiac mortality during long-term follow-up.
        Eur. Heart J.Cardiovasc. Imag. 2017; 18: 1331-1339
        • Han D.
        • et al.
        Prognostic significance of plaque location in non-obstructive coronary artery disease: from the CONFIRM registry.
        Eur. Heart J.Cardiovasc. Imag. 2022; 23: 1240-1247
      1. QIBA computed tomography angiography biomarkers committee.
        (QIBA Profile: Atherosclerosis Biomarkers by Computed Tomography Angiography (CTA)-2020 Profile Stage: Consensus. 2020; Available from:)
        • Sheahan M.
        • et al.
        Atherosclerotic plaque tissue: noninvasive quantitative assessment of characteristics with software-aided measurements from conventional CT angiography.
        Radiology. 2018; 286: 622-631
        • Stary H.C.
        Natural history and histological classification of atherosclerotic lesions: an update.
        Arterioscler. Thromb. Vasc. Biol. 2000; 20: 1177-1178
        • Stary H.C.
        • et al.
        A definition of advanced types of atherosclerotic lesions and a histological classification of atherosclerosis A report from the Committee on Vascular Lesions of the Council on Arteriosclerosis, American Heart Association.
        Circulation. 1995; 92: 1355-1374
        • Virmani R.
        • et al.
        Pathology of the vulnerable plaque.
        JACC (J. Am. Coll. Cardiol.). 2006; 47: C13-C18
        • Virmani R.
        • et al.
        Lessons from sudden coronary death: a comprehensive morphological classification scheme for atherosclerotic lesions.
        Arterioscler. Thromb. Vasc. Biol. 2000; 20: 1262-1275
        • Varga-Szemes A.
        • et al.
        Coronary plaque assessment of Vasodilative capacity by CT angiography effectively estimates fractional flow reserve.
        Int. J. Cardiol. 2021;
        • Zhu G.
        • et al.
        Semiautomated characterization of carotid artery plaque features from computed tomography angiography to predict atherosclerotic cardiovascular disease risk score.
        J. Comput. Assist. Tomogr. 2019;
        • Chrencik M.T.
        • et al.
        Quantitative assessment of carotid plaque morphology (geometry and tissue composition) using computed tomography angiography.
        J. Vasc. Surg. 2019;
        • Rafailidis V.
        • et al.
        Carotid Plaque Vulnerability: the Correlation of Plaque Components as Quantified Based on Computed Tomography Angiography with Neurologic Symptoms.
        European Congress of Radiology, 2019
        • Gupta A.
        • et al.
        Semi-automated detection of high-risk atherosclerotic carotid artery plaque features from computed tomography angiography.
        in: European Stroke Conference. 2017 (Berlin)
        • van Assen M.
        • et al.
        Automated plaque analysis for the prognostication of major adverse cardiac events.
        Eur. J. Radiol. 2019; 116: 76-83
        • Sigala F.
        • et al.
        Coronary versus carotid artery plaques. Similarities and differences regarding biomarkers morphology and prognosis.
        Curr. Opin. Pharmacol. 2018; 39: 9-18
        • Chatzizisis Y.S.
        • et al.
        Association of global and local low endothelial shear stress with high-risk plaque using intracoronary 3D optical coherence tomography: introduction of 'shear stress score'.
        Eur. Heart J.Cardiovasc. Imag. 2017; 18: 888-897
        • Gnasso A.
        • et al.
        In vivo association between low wall shear stress and plaque in subjects with asymmetrical carotid atherosclerosis.
        Stroke. 1997; 28: 993-998
      2. Ibrahimi, P., et al., Coronary and carotid atherosclerosis: how useful is the imaging? Atherosclerosis. 231(2): p. 323-333.

        • Schaar J.A.
        • et al.
        Terminology for high-risk and vulnerable coronary artery plaques. Report of a meeting on the vulnerable plaque, June 17 and 18, 2003, Santorini, Greece.
        Eur. Heart J. 2004; 25: 1077-1082