A Bayesian-Optimized Weighted Fuzzy Rank-Based Ensemble for Coronary Stenosis Detection in X-ray Angiography.
Authors
Affiliations (5)
Affiliations (5)
- Department of Artificial Intelligence, Faculty of Computer Engineering, University of Isfahan, Isfahan, Iran.
- Department of Artificial Intelligence, Faculty of Computer Engineering, University of Isfahan, Isfahan, Iran. [email protected].
- Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan, 84156-83111, Iran.
- Cardiac Rehabilitation Research Center, Cardiovascular Research Institute, Isfahan University of Medical Sciences, Isfahan, Iran.
- Interventional Cardiology Research Center, Cardiovascular Research Institute, Isfahan University of Medical Sciences, Isfahan, Iran.
Abstract
X-ray coronary angiography is the gold standard for diagnosing coronary artery disease, a leading cause of global mortality. Existing artificial intelligence methods typically detect stenosis from individual frames or rely on manually selected key frames, despite angiography being inherently a video-based examination in which diagnostically relevant information appears only in a subset of frames. Moreover, conventional ensemble methods generally employ fixed classifier weights, limiting their ability to account for uncertainty and variations in classifier performance. To address these challenges, we propose a framework that combines automatic key-frame selection with a novel weighted fuzzy rank-based ensemble model for right coronary artery (RCA) stenosis detection. The framework employs a semi-supervised key-frame selection strategy using ResNet50 or ViT-B/16 as feature extractors together with a hybrid loss function for vessel segmentation. The selected key frames are then classified using a weighted fuzzy rank-based ensemble, which extends conventional fuzzy rank-based fusion by incorporating learnable classifier weights optimized through Bayesian optimization. Video-level predictions are obtained via majority voting over key-frame classifications, enabling clinically meaningful sequence-level assessment. The proposed method was evaluated on the publicly available AngioCAD dataset and further validated on the independent CADICA dataset. Experimental results demonstrate a video-level precision of 84.63%, F1-score of 80.44%, ROC-AUC of 70.06%, and PR-AUC of 81.83% on AngioCAD, while maintaining consistent performance on the external dataset, indicating good generalizability. The proposed framework may assist cardiologists in the rapid and reliable identification of coronary stenosis and support clinical decision-making.