Improving Diabetic Retinopathy Detection Using Different Deep Learning Models with Data Balancing Techniques

Diabetic Retinopathy (DR) primarily affects people with persistent diabetes and is one of the world’s leading causes of blindness. Preventing vision loss requires early detection and precise classification of DR. In this study, we classify DR into five different categories: No DR, Mild DR, Moderate DR, Severe DR, and Proliferative DR. We do this by using three deep learning models: Convolutional Neural Network (CNN), MobileNet, and ResNet. We implemented data balancing approaches to provide more robust model training and evaluation because of the dataset’s intrinsic class imbalance. After a thorough evaluation of each model’s performance, the ResNet model had the best accuracy, coming in at 96%. Our results show that balanced data leads to a significant improvement in classification performance, which validates the effectiveness of the suggested method for automated, large-scale DR screening in clinical settings. The trade-offs between model accuracy and complexity are highlighted in this paper, which provides guidance on choosing suitable architectures for practical implementation.