Multi-Class Eye Diseases Prediction Using Ensemble CNN With Max Voting Strategy

Abstract

Eye diseases such as cataract, glaucoma, and diabetic retinopathy are major causes of blindness worldwide, underscoring the critical need for early and accurate diagnosis. This study presents a novel approach to multi-class eye disease prediction using an ensemble of Convolutional Neural Networks (CNNs) combined with a max voting strategy. The framework integrates four CNN models with varying architectural depths, each trained on a curated dataset of retinal fundus images, to classify eye conditions into four categories: cataract, glaucoma, diabetic retinopathy, and normal. The methodology begins with data pre-processing, which includes resizing, normalization, and augmentation to ensure robust model training. Each CNN model, ranging from six to nine layers, was trained independently for 60 epochs, leveraging techniques like dropout and regularization to prevent overfitting. The models' outputs were aggregated using a bagging ensemble technique, with final predictions determined through max voting. The ensemble approach effectively combines the complementary strengths of the individual models, enhancing classification reliability. Experimental evaluations demonstrate that the ensemble achieves superior performance, with an average accuracy of 92%, precision of 92%, recall of 91%, and F1-score of 92%. This study highlights the potential of combining deep learning with ensemble strategies for improved diagnostic accuracy in medical image analysis. By offering a scalable and reliable tool for early detection of eye diseases, this research contributes to advancing automated healthcare diagnostics, aiming to reduce the global burden of vision-related diseases and improve patient outcomes.