Topology-Aware Hierarchical Graph Vision Transformer with Structural Tokenization for Explainable Coronary Artery Stenosis Detection

Abstract

Coronary artery stenosis is a major cardiovascular condition that restricts blood flow due to abnormal vessel narrowing and requires accurate and timely diagnosis. However, automated stenosis detection from coronary angiography images remains challenging because of complex vascular topology, multi-scale vessel structures, and subtle morphological variations. Existing convolutional neural network and transformer-based approaches often fail to preserve vascular connectivity and hierarchical structural relationships, limiting detection accuracy and interpretability. To address these challenges, this study proposes a novel Graph-Guided Hierarchical Vision Transformer with Topology-Aware Structural Learning (GGH-ViT-TSL) framework for explainable coronary artery stenosis detection. The proposed framework integrates hierarchical CNN–ViT feature extraction, vessel segmentation, topology-preserving graph construction, and a newly designed Structural Tokenization mechanism that transforms vascular topology into graph-guided semantic tokens. These tokens are processed using a Graph Transformer to capture long-range inter-vessel dependencies and multi-scale stenotic characteristics, followed by an explainable detection module for stenosis classification and localization. The framework is implemented in Python and evaluated using the ARCADE coronary angiography dataset. Experimental results demonstrate that the proposed method achieves 98.2% accuracy, 98.1% precision, 98.3% recall, 98.2% F1-score, and an AUC of 0.990. In addition, vessel segmentation attains a Dice score of 0.965 and an IoU of 0.934. The proposed framework outperforms conventional CNN-based, graph-based, and transformer-based methods, confirming its effectiveness in preserving vascular topology and enhancing diagnostic interpretability. These findings demonstrate the potential of topology-aware graph-transformer learning for advanced medical image analysis and intelligent computer-aided cardiovascular diagnosis.