A Modified AlexNet-Based Deep Learning Framework for Multi-Disease Ophthalmic Classification and Severity Detection from Retinal Fundus Images

Abstract

Automated analysis of retinal fundus photographs has become an important research direction for computer-aided ophthalmic screening. However, many existing deep-learning studies focus on a single disease, particularly diabetic retinopathy, or emphasize disease presence without integrating severity assessment. This study presents a modified AlexNet convolutional neural network (CNN) framework for automated classification of ophthalmic disease from color fundus images, with a subsequent disease-specific severity assessment. The proposed framework uses contrast-limited adaptive histogram equalization (CLAHE), retinal-region segmentation, and morphological processing to improve the visibility of clinically relevant structures before deep feature learning. AlexNet is selected as the base network because of its relatively simple architecture and its ability to learn broad image structures with large initial kernels. Four architectural variants were investigated by changing network depth, input resolution, and the number of fully connected layers. The final architecture contains eleven convolutional layers, followed by two fully connected layers and a Softmax classifier. The experiments show that the 11-convolutional-layer configuration provides the best overall balance among feature extraction, classification performance, and computational complexity. In the reported experiments, the model achieved 80.86% accuracy and 88.89% specificity. These results are used as a base to select the variant for further research work. Then the final selected variant with an 11x11 convolutional-layer framework was tested on benchmark datasets MESSIDOR, Drishti-GS, APTOS, and a real-time clinical dataset and achieved remarkable accuracy for ADAM, NDAM, RMsprop, SGD, Adagrad, Adamax, and Adadelta optimizers. The framework further represents diabetic retinopathy severity using normal, mild NPDR, moderate NPDR, severe NPDR, and proliferative DR categories, and glaucoma severity using normal, early, mild, moderate, and severe categories. The study contributes an integrated fundus-image analysis pipeline that combines multi-scale deep feature learning with disease-specific severity assessment and provides a reproducible basis for further clinical validation.