Coronary motion as a functional biomarker in routine X-ray angiography

Background Cardiovascular disease remains a leading cause of mortality worldwide, and a considerable proportion of patients with coronary artery disease present without traditional risk factors, highlighting the need for additional functional biomarkers. Coronary artery motion observed in routine X-ray angiography reflects myocardial contraction and vascular mechanics but remains underexplored as a quantitative biomarker due to challenges in robust and scalable motion extraction. Purpose This study aimed to develop and validate a scalable deep learning framework to extract coronary artery motion from routine angiographic sequences and to evaluate associations between motion-derived representations and established cardiovascular risk factors. Methods In this retrospective study, >50,000 angiographic sequences from 1,824 patients were analysed. To reduce geometric variability, analyses were restricted to the dominant projection cluster, yielding 8,792 sequences. Nine consecutive peak-contrast frames per sequence were selected. Coronary motion between consecutive frames was quantified using a deep learning–based dense optical flow method and validated using synthetic transformations and reconstruction-based assessment on real data. Stacked motion fields were used to train a network capturing spatial and temporal coronary dynamics. Two video-based architectures were evaluated (a two-stream ResNet-152 and a transformer-based model), with the ResNet-152 demonstrating superior performance and selected for subsequent analyses. The network was trained to predict categorical clinical parameters. Results On synthetic data, motion estimation achieved a mean endpoint error of 0.50 pixels with 96.1% of displacements within one pixel. On real angiographic data, motion-based reconstruction reduced mean squared error from 1944.9 to 1710.6 and increased sensitivity from 0.49 to 0.55 compared with a no-motion baseline. Motion representations achieved higher accuracy for reduced ejection fraction (83.0% vs 53%, p<0.001), hypercholesterolaemia (68.4% vs 42%, p<0.001), hypertension (90.2% vs 68%, p<0.01), and nicotine use (81.5% vs 54%, p<0.001), with no significant improvement for diabetes. Conclusion Coronary artery motion can be extracted from routine X-ray angiography in a scalable, segmentation-free manner. Motion-derived representations demonstrate significant associations with established cardiovascular risk factors, supporting coronary motion as a promising functional imaging biomarker.