A self-supervised hybrid CNN with uncertainty-aware referral for diabetic retinopathy screening

Diabetic retinopathy (DR) is a leading cause of vision loss, making accurate, early screening essential. We present a lightweight, self-supervised hybrid CNN for automated DR detection from retinal fundus images that fuses MobileNet and EfficientNet with a residual attention mechanism for multi-scale, DR-specific feature extraction, and combines predictions in an ensemble with ResNet for robustness. The training pipeline includes a two-stage self-supervised warm start (general EyePACS pretraining followed by domain-specific refinement), a binary-to-five-class curriculum schedule to stabilize optimization, Grad-CAM for visual interpretability, and an uncertainty-aware referral module based on Monte Carlo Dropout. Evaluated on public (APTOS, Messidor) and private datasets under stratified 5-fold cross-validation with strict patient-level separation and external held-out testing on an independent private dataset, the framework achieves state-of-the-art performance with a macro F1-score of 98.1% and an AUC of 99.2% (mean across test folds), representing a 2.1 percentage-point gain over an ImageNet-initialized baseline. The uncertainty module flags 75% of misclassified cases for expert review, while a fairness audit stratified by image quality shows stable performance with a worst-case AUC drop of only 1.2 percentage points. Inference is efficient (approximately 45 ms on a consumer GPU for the full ensemble and approximately 180 ms on a mobile-class CPU for a single forward pass of the hybrid backbone), supporting deployment from clinic to edge. These results indicate that a self-supervised hybrid CNN with explainability and uncertainty-aware referral can deliver accurate, reliable, and equitable DR screening with promising potential for real-world clinical workflows.