DIABETIC RETINOPATHY DETECTION USING CONVOLUTIONAL NEURAL NETWORKS: A DEEP LEARNING APPROACH

Abstract

Diabetic retinopathy (DR) is a progressive eye disease and a leading cause of preventable blindness among diabetic patients, especially in regions with limited access to specialized healthcare. This thesis presents the design, implementation, and evaluation of a Convolutional Neural Network (CNN)-based system for automated detection and classification of diabetic retinopathy using the APTOS 2019 dataset. The project addresses the challenge of class imbalance across five DR stages—No DR, Mild, Moderate, Severe, and Proliferative DR—by employing targeted data augmentation and careful preprocessing, including image normalization and contrast enhancement. The proposed CNN architecture, featuring four convolutional layers and dropout regularization, was trained and validated on stratified splits of the dataset, achieving a test accuracy of 74.28%. The model demonstrated high precision and recall for "No DR" cases and reasonable performance on intermediate and advanced stages, as evidenced by confusion matrices, ROC curves, and classification reports. Explainable AI techniques, such as Grad-CAM and LIME, were integrated to visualize the regions and features that influenced the model’s predictions, ensuring alignment with clinically relevant lesions and supporting transparency for clinical adoption. The lightweight and efficient design of the model allows for real-time inference on standard hardware, making it suitable for deployment in community health programs and telemedicine platforms. Overall, this work demonstrates the feasibility and effectiveness of deep learning for early DR detection, and provides a foundation for future improvements in automated ophthalmic diagnostics and large scale screening initiatives.