Abstract

Early detection of ocular diseases is critical for preventing irreversible vision loss. We propose SGAP-Net, a lightweight deep learning model for classifying four retinal conditions, cataract, diabetic retinopathy, glaucoma, and normal, from fundus images. Built on EfficientNet-B3, SGAP-Net incorporates a Saliency-Guided Attention Path (SGAP) module that enhances focus on diagnostically relevant regions with minimal computational overhead (1.02 G FLOPs, 10.62 M parameters). Key improvements over the original version include Contrast Limited Adaptive Histogram Equalization (CLAHE), structure-preserving augmentations (e.g., Random Erasing, rotation), label smoothing, and class-balanced loss, collectively boosting accuracy to 88.15% and AUC-ROC to 97.33% on a stratified test set. While sensitivity for the normal class remains moderate (79.18%), indicating a tendency toward false positives, quantitative and visual analyses confirm that model attention consistently aligns with clinical biomarkers, particularly retinal vasculature and the optic disc. The system also enables automated cup-to-disc ratio (CDR) estimation. Ablation studies demonstrate that these gains stem from training refinements rather than architectural complexity. SGAP-Net thus offers a favorable trade-off between accuracy, efficiency, and interpretability, underscoring the value of careful preprocessing, augmentation, and loss design in resource-constrained retinal screening.

Concepts :

Access to Document

Citations by Year