Quantifying regional contributions to intercourse classification from fundus images by way of a two-stage attention-based deep studying strategy

This research quantitatively assesses the relative contribution of anatomically outlined retinal areas (macula, optic disc, and retinal vasculature) to organic intercourse classification from fundus images. We developed a two-stage multi-branch framework utilizing pre-trained ResNet50 backbones for region-specific characteristic extraction and an attention-based fusion module that produces ROI-level scalar weights. The dataset comprised 3,478 eye-level fundus photos with a balanced intercourse distribution, and knowledge had been cut up on a subject-wise foundation to keep away from info leakage between units. The fusion mannequin achieved the very best discriminative efficiency (AUC = 0.861; 95% CI: 0.826–0.895). Single-branch outcomes recommended complementary tendencies, with the macula department displaying larger sensitivity for feminine classification and the optic disc department displaying larger specificity for male classification. In the fusion mannequin, imply consideration weights for the macula (0.408; 95% CI: 0.369–0.446) and optic disc (0.383; 95% CI: 0.347–0.418) had been comparable (p = 0.506), and each had been considerably larger than these for the vasculature (0.209; 95% CI: 0.184–0.238; p < 0.001). These region-level weights present a reproducible, quantitative estimate of mannequin reliance throughout predefined ROIs and complement qualitative explainability strategies. Overall, the proposed framework helps extra clear and anatomically interpretable AI for fundus-based intercourse classification.