A Framework for Cross-Domain Generalization in Coronary Artery Calcium Scoring Across Gated and Non-Gated Computed Tomography

Mahmut S. Gokmen; Moneera N. Haque; Steve W. Leung; Caroline N. Leach; Seth Parker; Stephen B. Hobbs; Vincent L. Sorrell; W. Brent Seales

cs.CVFeb 25, 2026

A Framework for Cross-Domain Generalization in Coronary Artery Calcium Scoring Across Gated and Non-Gated Computed Tomography

Authors: Mahmut S. Gokmen, Moneera N. Haque, Steve W. Leung, Caroline N. Leach, Seth Parker, Stephen B. Hobbs, Vincent L. Sorrell, W. Brent Seales,

Organizations: Center for Applied AI, University of Kentucky, Lexington, KY · Department of Medicine, Division of Cardiovascular Medicine and the Gill Heart and Vascular Institute, University of Kentucky, Lexington, KY · Department of Radiology, College of Medicine, University of Kentucky, Lexington, KY · EduceLab, University of Kentucky, Lexington, KY

Abstract

Coronary artery calcium (CAC) scoring is a key predictor of cardiovascular risk, but it relies on ECG-gated CT scans, restricting its use to specialized cardiac imaging settings. We introduce an automated framework for CAC detection and lesion-specific Agatston scoring that operates across both gated and non-gated CT scans. At its core is CARD-ViT, a self-supervised Vision Transformer trained exclusively on gated CT data using DINO. Without any non-gated training data, our framework achieves 0.707 accuracy and a Cohen's kappa of 0.528 on the Stanford non-gated dataset, matching models trained directly on non-gated scans. On gated test sets, the framework achieves 0.910 accuracy with Cohen's kappa scores of 0.871 and 0.874 across independent datasets, demonstrating robust risk stratification. These results demonstrate the feasibility of cross-domain CAC scoring from gated to non-gated domains, supporting scalable cardiovascular screening in routine chest imaging without additional scans or annotations.

Figures & tables

Figure 1: Overview of the automated CAC scoring architecture for coronary artery calcium detection and quantification.

Panel of Figure 2 — Figure 2: Attention maps from CARD-ViT highlighting calcified regions in non-gated (a) and gated (b) CT scans. Each pair shows a raw CT slice (left) and the corresponding PCA-transformed attention visualization (right) from the self-supervised ViT backbone.

Figure 3: CARD-ViT segmentation pipeline: ViT feature extraction, $784$ -channel feature fusion, decoder upsampling with selective skip connections, and final binary mask output.

Figure 4: Automated segmentation by CARD-ViT (red overlay) visualized in the OHIF Viewer via MONAI Label for real-time clinical review.

	Case-Level			Slice-Level			Available		CAC Score Distribution
Dataset / Vessel	Total	Labeled	%	Total	Labeled	%	Score	Label	0–10	11–100	101–400	400+
Dataset-Level Statistics
Heartlens (Gated Train)	2,651	727	27.4	162,584	7,028	4.3	✓	✓	1,220	504	456	471
Heartlens (Gated Test)	469	129	27.5	28,873	1,209	4.2	✓	✓	216	89	81	83
Stanford (Gated Test)	443	443	100.0	22,695	3,602	15.9	✗*	✓	77	140	103	123
Stanford (Non-Gated)	205	–	–	9,222	–	–	✓	✗	105	42	32	27

Table 1: Comprehensive dataset statistics. Top section shows dataset-level case/slice counts, annotation coverage, and CAC score distribution. Bottom section shows vessel-wise annotation breakdown for Heartlens gated cohort (all splits combined). ”Labeled” indicates availability of pixel-level annotations. Percentages in case/slice columns indicate annotation coverage. ”Score” indicates patient-level CAC score availability; ”Label” indicates pixel-level segmentation label availability. For the Stanford (Gated Test) cohort, the CAC score distribution was derived from the provided pixel-level ground-truth (GT) annotations, as patient-level scores were unavailable.

Dataset	N (Total Slices)	Dice	IoU	Precision	Recall
Heartlens Test	28,873	0.865	0.857	0.869	0.867
Stanford Test	22,695	0.839	0.829	0.844	0.845

Table 2: Segmentation performance on annotated slices from gated test sets. Metrics computed only on slices with ground-truth annotations.

N

indicates the total number of test slices; metrics are averaged over annotated slices within each test set.

Panel of Figure 5 — Figure 5: Qualitative segmentation comparison on gated CT slices. Left: Ground-truth annotation overlay. Middle: CARD-ViT prediction showing accurate calcification localization. Right: AI-CAC prediction (trained on non-gated data) demonstrating domain mismatch with over-segmentation artifacts.

Test Cohort	Model	Accuracy	Cohen’s $\kappa$
Heartlens (Gated)	CARD-ViT	0.910	0.871
Stanford (Gated)	CARD-ViT	0.910	0.874
Stanford (Non-Gated)	CARD-ViT	0.707	0.528
Stanford (Non-Gated)	AI-CAC	0.707	0.542

Table 3: Summary of CAC scoring performance across gated and non-gated test cohorts. CARD-ViT was trained exclusively on gated data; non-gated evaluation represents cross-domain transfer. AI-CAC 11 serves as a baseline trained directly on non-gated scans.

CAC Range	Heartlens					Stanford
CAC Range	Sens	Spec	PPV	NPV	F1	Sens	Spec	PPV	NPV	F1
0–10	0.93	0.99	0.98	0.96	0.95	0.82	0.97	0.94	0.89	0.88
11–100	0.88	0.97	0.84	0.98	0.86	0.90	0.96	0.91	0.97	0.90
101–400	0.86	0.97	0.84	0.98	0.85	0.91	0.97	0.85	0.98	0.88
400+	0.95	0.99	0.91	0.99	0.94	0.97	0.98	0.92	0.99	0.95
Overall (Acc/ $\kappa$ )	0.910 / 0.871					0.910 / 0.874

Table 4: Gated CT performance on Heartlens and Stanford test sets. Top section shows per-category performance metrics; bottom section displays confusion matrices with ground truth in rows and predictions in columns.

CAC Range	CARD-ViT (Gated $\rightarrow$ Non-Gated)					AI-CAC (Trained on Non-Gated)
CAC Range	Sens	Spec	PPV	NPV	F1	Sens	Spec	PPV	NPV	F1
0–10	0.933	0.720	0.778	0.911	0.848	0.876	0.770	0.800	0.856	0.836
11–100	0.415	0.896	0.500	0.860	0.453	0.463	0.890	0.514	0.869	0.487
101–400	0.438	0.942	0.583	0.901	0.500	0.406	0.936	0.542	0.895	0.464
400+	0.593	0.972	0.762	0.940	0.667	0.778	0.955	0.724	0.966	0.750
Overall (Acc/ $\kappa$ )	0.707 / 0.528					0.707 / 0.542

Table 5: Cross-domain generalization performance on non-gated CT scans, comparing CARD-ViT (trained only on gated data) with AI-CAC (trained directly on non-gated data) on the Stanford non-gated test set. Top section shows per-category metrics; bottom section displays confusion matrices with ground truth in rows and predictions in columns.

Backbone	Dice	IoU	Precision	Recall
DINOv2-with-registers-Small	0.672	0.651	0.667	0.675
DINOv2-with-registers-Base	0.865	0.857	0.869	0.867
DINOv2-with-registers-Large	0.835	0.823	0.842	0.832
DINOv2-with-registers-Giant	0.801	0.788	0.814	0.790

Table 6: Ablation results comparing DINOv2 backbone sizes (Small to Giant) on the gated Heartlens test set. Performance metrics include Dice, IoU, Precision, and Recall.

Domain Generalization

Domain generalization develops models that perform reliably in new environments without training on target-domain data.

See all 375 papers

Quantitative coronary calcification analysis for prediction of myocardial ischemia using non-contrast CT calcium scoring
May 20, 2026Juhwan Lee, Sadeer Al-Kindi, Ammar Hoori +6Clinical Outcome Prediction Radiomics
Department of Biomedical Engineering, Virginia Commonwealth University, Richmond, VA, 23284, USA · Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, 44106, USA · DeBakey Heart and Vascular Institute, Houston Methodist Hospital, Houston, TX, 77030 +2
CORA: Generalizable coronary artery disease assessment and risk stratification from coronary CT angiography using pathology-centric representation learning
Mar 25, 2026Jinkui Hao, Gorkem Durak, Halil Ertugrul Aktas +4Synthetic Data Pretraining Coronary Angiography
Department of Radiology, Northwestern University, Chicago, IL, USA · Department of Biomedical Engineering, Northwestern University, Evanston, IL, USA · Department of Cardiology, Northwestern University, Chicago, IL, USA +1
CARDINAL Predicts Cardiovascular Risk From Non-contrast Cardiac CT
Aug 27, 2026Roy Gabriel, Nattakorn Kittisut, Jamshid Hassanpour +6Cardiovascular Risk Prediction Clinical Outcome Prediction
Georgia Institute of Technology, Atlanta, Georgia, USA · Emory University Hospital, Atlanta, Georgia, USA

May 20, 2026J/K move · Enter open · S save

Quantitative coronary calcification analysis for prediction of myocardial ischemia using non-contrast CT calcium scoring

Juhwan Lee, Sadeer Al-Kindi, Ammar Hoori, Tao Hu+5

Department of Biomedical Engineering, Virginia Commonwealth University, Richmond, VA, 23284, USA · Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, 44106, USA · DeBakey Heart and Vascular Institute, Houston Methodist Hospital, Houston, TX, 77030+2

Clinical Outcome Prediction Radiomics Computed Tomography Clinical Risk Prediction Cardiovascular Risk Prediction

Open paper AskPDF arXiv

Non-contrast computed tomography calcium scoring (CTCS) is widely recognized as an effective tool for cardiovascular risk stratification. This study aimed to develop a novel machine learning framework for predicting myocardial ischemia from routine non-contrast CTCS scans using quantitative coronary calcium assessment. This study analyzed 1,375 patients who underwent both non-contrast CTCS and regadenoson stress cardiac positron emission tomography myocardial perfusion imaging within one year at University Hospitals Cleveland Medical Center. A total of 74 variables, including clinical variables, Agatston score, and calcium-omics features, were evaluated. Relevant features were identified using XGBoost with Shapley Additive exPlanations (SHAP). Predictive models were trained and evaluated using 5-fold cross-validation. Among 987 patients, 89 (9%) were positive for myocardial ischemia. The final model incorporated the Agatston score, eight calcium-omics features, and age. The proposed model achieved a precision of 98.9+/-3.0%, sensitivity of 79.2+/-8.4, and F1 score of 87.7+/-5.3%. The addition of calcium-omics features significantly improved predictive performance compared with models using clinical variables alone or clinical variables with the Agatston score (p<0.05). Interestingly, the number of calcified arteries, despite being the lowest-ranked feature based on SHAP analysis, showed the strongest association with myocardial ischemia in logistic regression analysis (odds ratio: 3.63, 95% confidence interval: 2.80-4.77, p<0.00001). We developed a machine learning approach for predicting myocardial ischemia using routinely acquired non-contrast CTCS scans. Calcium-omics features provided incremental predictive value beyond conventional risk factors and Agatston scoring and may support more accessible cardiovascular risk stratification.

Abstract

Domain Generalization

Explore similar work

Quantitative coronary calcification analysis for prediction of myocardial ischemia using non-contrast CT calcium scoring

CORA: Generalizable coronary artery disease assessment and risk stratification from coronary CT angiography using pathology-centric representation learning

CARDINAL Predicts Cardiovascular Risk From Non-contrast Cardiac CT