Coronary Artery Classification Using Deep Learning

Résumé fourni par la source

Automated interpretation of coronary angiography has the potential to improve diagnostic efficiency and reduce intra- and inter-observer variability in the assessment of coronary artery disease. A critical early step in this process is the classification of angiographic sequences into left coronary artery (LCA) and right coronary artery (RCA) views. In this study, we evaluate three deep learning models: ResNet50, Swin Transformer (Swin-B), and Vision Transformer (ViT-B) for single-frame-based LCA/RCA classification. We use a majority-voting scheme to aggregate framelevel predictions into robust video-level classifications. Our dataset comprises 45,901 frames from 2,499 angiographic videos collected from a single-center cohort of 376 patients. ResNet50, trained from scratch on our domain-specific data, consistently outperformed both transformer-based models, which were pretrained on natural images. At video-level, ResNet50 achieved an F1-score of 0.9897, followed by ViT-B (0.9475) and Swin-B (0.9225). The transformer models were more sensitive to class imbalance, resulting in lower specificity. Applying majority voting significantly improved all models' performance by an average of 1.98 %. These results demonstrate that single-frame-based classification, combined with temporal aggregation, can achieve high accuracy while retaining flexibility for downstream tasks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.