Early CHD Detection from Retinal Fundus Scans Using a Spatial Context-Aware Hierarchical Attention Framework

Sparsh Rastogi, Khushboo Modi, Sahil Thakur, Vinay Arora

Abstract: Retinal fundus imaging provides a non-invasive, cost-effective, and scalable modality for early detection of systemic diseases such as coronary heart disease (CHD). This study proposes a Swin Transformer-based deep learning framework for CHD classification from retinal fundus images, enhanced through vascular segmentation and model interpretability. A total of 17,242 images were curated from the UK Biobank, with class imbalance mitigated using optimal transport-based sampling. Retinal vasculature was segmented using LW-Net to extract both full vessel maps and isolated arteriolar structures, which served as anatomically enriched inputs to the Swin Transformer classifier. The model achieved an AUROC of 0.81 using raw fundus images, which improved to 0.90 with vessel maps and 0.87 with arteriole-only maps—highlighting the benefit of vessel-focused preprocessing. Grad-CAM visualizations revealed consistent attention around the optic disc and major arterioles, reinforcing clinical relevance and model transparency. These findings establish a robust and interpretable pipeline for CHD risk prediction and support the broader utility of retinal imaging for cardiovascular screening, especially in resource-constrained settings.

External IDs:dblp:conf/miccai/RastogiMTA25