Deep Learning-Based Ophthalmic Disease Classification With Ensemble Modeling And Grad-CAM Visualization

Abstract

Cataract, diabetic retinopathy (DR), and glaucoma are ophthalmic diseases that cause vision impairment and blindness worldwide. For safeguardingtimely treatment and preventing irreversible vision loss, early detection and diagnosis are vital. This study presents a comprehensive and interpretable deep learning framework that blends multiple pre-trained convolutional neural network architectures, ensemble learning, and Grad-CAM for accurate and transparent segmentation of retinal fundus imagesfor multi-class classification of ophthalmic diseases. Four CNN models, i.e. ResNet-50, InceptionV3, GoogLeNet, and AlexNet were adjusted through transfer learning approaches and trained on a curated dataset comprising four categories: Cataract, Diabetic Retinopathy, Glaucoma, and Normal. Data augmentation approaches, such as rotation, translation, and scaling, were employed to enhance model robustness and generalization. Each model was evaluated in context toperformance evaluation parameters such as accuracy, confusion matrices, and ROC curves. Further, an ensemble model was constructed to combine the strengths of individual networks, achieving improved diagnostic accuracy. ResNet-50 attained the highest classification accuracy of 93.80 % among the individual model, while the ensemble approach demonstrated superior overall performance with accuracy of 94.14 %. To address the limitations caused by opaque decision-making process of CNN, visualization methods such as Gradient-weighted Class Activation Mapping (Grad-CAM) are employed to interpret class activation regions and validate model focus on clinically relevant areas to boost interpretability. The proposed framework establishes a promising foundation for clinical decision support systems and AI-assisted classification of ophthalmic diseases that not only achieves high diagnostic reliability but also bridges performance and explainability.