A dual-stage deep learning framework for detecting borderline coronary artery stenosis in x-ray angiography
Year: 2026
Type: article
Abstract: Coronary artery disease (CAD) is a leading cause of mortality worldwide. The early identification of coronary artery stenosis using x-ray angiography images can help to prevent myocardial infarction. Existing methods have issues identifying thin and indistinct vessels, and they fail to detect long vessel segments. To overcome these limitations, this study proposes a dual-stage deep learning framework that integrates segmentation and detection to enhance the automatic detection of borderline patients using x-ray angiography images. Here, we used a novel residual attention U-net plus (RA-UNet++) by embedding a dual-attention mechanism (channel and spatial attention) within a residual U-net ++ encoder-decoder to segment the vessels and stenosis regions and employing attention mechanisms to emphasize subtle and narrow vascular features. After segmentation, a YOLOv8 model was used to detect the stenotic region and identify patients with stenosis levels of 40%–70%, thereby improving diagnostic specificity. An experimental evaluation of the automatic region-based CAD diagnostics using x-ray angiography images dataset demonstrated that the proposed segmentation model achieved an F 1-score of 0.8384, Dice coefficient of 0.8318, precision of 0.8612, and recall of 0.8100. The detection stage achieves an F 1-score of 0.60 for stenosis identification. These results indicate that the proposed framework provides an effective automated approach to assist clinicians in detecting borderline coronary artery stenosis.
Source: Engineering Research Express
Authors: C K Revathi, H Santhi
Institution: Vellore Institute of Technology University
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,400
OpenAlex ID: W7155386302