RetinalNet: Nine-Class Retinal Disease Classification Using EfficientNet-B2 on Multi-Dataset Fundus Images

Background: Diabetic retinopathy (DR), glaucoma, age-related macular degeneration (AMD), and cataracts cause most preventable blindness, yet screening coverage is lowest where ophthalmologists are scarcest. Existing deep learning systems are condition-specific, requiring separate models per disease. We developed and internally validated RetinalNet, a unified nine-class classifier assigning a single fundus photograph to one of five DR grades (no DR through proliferative), glaucoma, cataracts, AMD, or other ocular diseases. Normal images fall within the no-DR grade, making this a nine-class rather than ten-class task. Methods: Three public datasets (APTOS-2019, ODIR-5K, RIM-ONE DL) were harmonized into a unified nine-class taxonomy of 15,729 images. The corpus was severely imbalanced (training class sizes 136–5,100; ratio 37.5:1; two largest classes 66.4% of training data). ImageNet-pretrained EfficientNet-B2 was fine-tuned end-to-end with focal loss, inverse-frequency class weighting, AdamW, and mixed precision. A held-out split (n=1,887) was evaluated for accuracy, per-class precision, recall and F1, ROC-AUC, and precision–recall AUC (PR-AUC) referenced to class prevalence; Grad-CAM assessed anatomical fidelity; the model was exported to ONNX. Results: Accuracy was 91.79% (macro-F1 0.832; weighted-F1 0.917). ROC-AUC exceeded 0.967 for all nine classes; macro-averaged PR-AUC was 0.856 (weighted 0.949). Cataracts (F1 0.990), AMD (0.956), and glaucoma (0.919) performed best, each with PR-AUC above 0.98. The rare DR grades were weakest: Mild and Proliferative DR reached F1 0.636 and 0.533, while Severe DR showed the lowest PR-AUC (0.571) despite an ROC-AUC of 0.982, reflecting prevalence under 2% and as few as 18 validation cases. Errors concentrated within the DR severity continuum; none occurred between structurally distinct diseases. Grad-CAM localized to the optic disc for glaucoma and the macula for AMD and DR. CPU inference ran under 80 ms. Conclusions: A single lightweight model can simultaneously screen for the leading causes of preventable blindness on standard hardware. Validation was internal; external clinical validation is required before deployment.