Quantifying regional contributions to sex classification from fundus photographs via a two-stage attention-based deep learning approach.

This study quantitatively assesses the relative contribution of anatomically defined retinal regions (macula, optic disc, and retinal vasculature) to biological sex classification from fundus photographs. We developed a two-stage multi-branch framework using pre-trained ResNet50 backbones for region-specific feature extraction and an attention-based fusion module that produces ROI-level scalar weights. The dataset comprised 3,478 eye-level fundus images with a balanced sex distribution, and data were split on a subject-wise basis to avoid information leakage between sets. The fusion model achieved the highest discriminative performance (AUC = 0.861; 95% CI: 0.826-0.895). Single-branch results suggested complementary tendencies, with the macula branch showing higher sensitivity for female classification and the optic disc branch showing higher specificity for male classification. In the fusion model, mean attention weights for the macula (0.408; 95% CI: 0.369-0.446) and optic disc (0.383; 95% CI: 0.347-0.418) were comparable (p = 0.506), and both were significantly higher than those for the vasculature (0.209; 95% CI: 0.184-0.238; p < 0.001). These region-level weights provide a reproducible, quantitative estimate of model reliance across predefined ROIs and complement qualitative explainability methods. Overall, the proposed framework supports more transparent and anatomically interpretable AI for fundus-based sex classification.