Model performance summary.

Diabetic Retinopathy, Cataract, and Glaucoma are major retinal diseases that require early detection to prevent irreversible vision loss. This study proposes a deep learning-based framework for the automated classification of retinal images into four categories: Normal, Diabetic Retinopathy, Cataract, and Glaucoma. The dataset was compiled from publicly available retinal imaging databases, including IDRiD and HRF. Four convolutional neural network architectures—EfficientNet-B0, EfficientNet-B7, a build-from-scratch model, and AlexNet—were evaluated using multiple performance metrics. Among these, AlexNet achieved the highest overall performance, attaining an accuracy of 93.65%, sensitivity of 94.39%, specificity of 98.05%, PPV of 93.65%, NPV of 97.95%, and an F1-score of 93.74%. EfficientNet-B7 followed with an accuracy of 92.82%, confirming the strength of transfer learning in retinal feature extraction. A five-fold cross-validation further validated AlexNet’s robustness, yielding a mean R2 of 0.8891 with low variance, indicating consistent generalization across folds. Computational efficiency analysis showed that AlexNet achieved high diagnostic accuracy with a moderate processing time of approximately 14 minutes. Model interpretability using SHapley Additive exPlanations (SHAP) revealed that AlexNet highlighted clinically relevant retinal regions, such as the optic disc and macula, thereby enhancing transparency and clinical trust. In summary, the proposed framework demonstrates that interpretable deep learning models can deliver accurate, consistent, and explainable retinal disease classification, offering a foundation for real-time, AI-assisted ophthalmic screening systems.