Multiclass Eye Disease Recognition And Classification Using An Ensemble Deep Transfer Learning Framework

Abstract

Early detection of eye diseases is vital to prevent severe vision impairment and blindness, as many ocular disorders progress silently neural networks, MobileNetV2, VGG19, and VGG16, to enhance classification accuracy, robustness, and generalizability in multiclass eye disease recognition. A publicly available Kaggle dataset comprising 383 retinal images across five categories (Glaucoma, Cataracts, Bulging Eyes, Crossed Eyes, and Uveitis) was utilized, and data augmentation techniques expanded the dataset to 9,000 images to reduce overfitting and improve model generalization. The proposed ensemble achieved a classification accuracy of 97.93%, out per forming individual architectures and benchmark models such as ResNet50, DenseNet121, and EfficientNetB0, thereby demonstrating a strong until advanced stages. In this study, we propose an ensemble deep learning framework that integrates three pre-trained convolutional balance between predictive performance and computational efficiency. The findings underscore the potential of ensemble CNN to support ophthalmologists in reliable and timely screening, ultimately reducing the burden of preventable blindness. Future directions include integrating explainable AI approaches, such as Grad-CAM and attention mechanisms, to improve interpretability and clinician trust, as well as extending the framework through semi-supervised learning for rare or co-occurring ocular pathologies.