CORA: A Pathology Synthesis Driven Foundation Model for Coronary CT Angiography Analysis and MACE Risk Assessment

Abstract

Coronary artery disease, the leading cause of cardiovascular mortality worldwide, can be assessed non-invasively by coronary computed tomography angiography (CCTA). Despite progress in automated CCTA analysis using deep learning, clinical translation is constrained by the scarcity of expert-annotated datasets. Furthermore, widely adopted label-free pretraining strategies, such as masked image modeling, are intrinsically biased toward global anatomical statistics, frequently failing to capture the spatially localized pathological features of coronary plaques. Here, we introduce CORA, a 3D vision foundation model for comprehensive cardiovascular risk assessment. CORA learns directly from volumetric CCTA via a pathology-centric, synthesis-driven self-supervised framework. By utilizing an anatomy-guided lesion synthesis engine, the model is explicitly trained to detect simulated vascular abnormalities, biasing representation learning toward clinically relevant disease features rather than dominant background anatomy. We trained CORA on a large-scale cohort of 12,801 unlabeled CCTA volumes and comprehensively evaluated the model across multi-center datasets from nine independent hospitals. Across diagnostic and anatomical tasks, including plaque characterization, stenosis detection, and coronary artery segmentation, CORA consistently outperformed the state-of-the-art 3D vision foundation models, achieving up to a 29\% performance gain. Crucially, by coupling the imaging encoder with a large language model, we extended CORA into a multimodal framework that significantly improved 30-day major adverse cardiac event (MACE) risk stratification. Our results establish CORA as a scalable and extensible foundation for unified anatomical assessment and cardiovascular risk prediction.

Figures (10)

• Figure 1: Overview of CORA and synthesis-driven pretraining framework. (a) Synthesis-driven self-supervised pretraining of CORA on large-scale unlabeled CCTA volumes. An anatomy-guided lesion synthesis engine generates diverse calcified and non-calcified plaque patterns with controlled morphology and attenuation, which are inserted into CCTA volumes to create simulated abnormality. Multi-windowing inputs capture complementary tissue characteristics across soft tissue, contrast-enhanced lumen, and calcium windows, enabling the encoder to learn pathology-centric representations and produce localized abnormality response maps. (b) Down-stream clinical assessment using the pretrained CORA encoder. A single foundation model supports multiple downstream tasks, including coronary plaque classification, stenosis detection, and coronary artery segmentation. For prognostic modeling, imaging representations are further integrated with patient demographic and clinical information through a multimodal fusion framework to enable major adverse cardiac event (MACE) risk stratification.
• Figure 2: Performance of plaque characterization across internal and external cohorts. Receiver operating characteristic (ROC) curves for the detection of calcified and non-calcified coronary plaques on the internal test cohort (a,b) and external multi-center cohort (c,d). CORA consistently outperformed models trained from scratch and existing 3D foundation models, with the largest performance gains observed on the external cohort, indicating superior generalization under domain shift. e, Visualization of model attention using Grad-CAM showed model accurately identified plaque regions in CCTA volumes. The 95% CIs for AUCs were computed using 500 bootstrap resamples. Differences in AUCs were assessed using a bootstrap test. Statistical significance is denoted as $*p < 0.05$, $**p < 0.01$, and $***p < 0.001$.
• Figure 3: CORA performance in coronary stenosis detection. (a) Representative axial CCTA images illustrating the target pathologies. Magnified insets detail a low-contrast non-calcified stenosis (left, blue) and a calcified stenosis (right, red). (b) Distribution of annotated stenoses in the entire fine-tuning dataset, stratified by CAD-RADS severity grade and plaque composition (calcified, red; non-calcified, blue). (c) Quantitative benchmarking against a baseline (From Scratch) and self-supervised 3D foundation models (MAE, VolumeFusion, VoCo). CORA (red) achieves state-of-the-art performance across Precision, Recall, and F1 Score.
• Figure 4: Coronary artery segmentation performance and data efficiency. (a) Segmentation performance on the ImageCAS test set using 100 labeled training samples, evaluated by Dice similarity coefficient, centerline Dice (clDice), and mean surface distance (MSD). CORA outperformed models trained from scratch and existing 3D foundation models across all metrics. (b) Data efficiency analysis across varying numbers of labeled training samples (50–800). CORA consistently achieved higher segmentation accuracy and better centerline preservation than the baseline, with the largest gains observed in low-data regimes.
• Figure 5: CORA performance in MACE prediction. Receiver operating characteristic (ROC) curves for MACE risk stratification on the internal (a) and external (b) validation datasets. The vision-only CORA(Image) (light red) consistently outperforms the supervised baseline and state-of-the-art 3D foundation models (MAE, VolumeFusion, VoCo). Fusing clinical metadata with imaging representations, CORA(Multi-modal) (solid red) achieves the highest prognostic accuracy and robust cross-center generalization. The 95% CIs for AUCs were computed using 500 bootstrap resamples. Differences in AUCs were assessed using a bootstrap test. Statistical significance is denoted as $*p < 0.05$, $**p < 0.01$, and $***p < 0.001$.