Enhanced Diabetic Retinopathy Detection An Explainable SemiSupervised Approach Using Contrastive Learn

Diabetic Retinopathy (DR), a leading cause of blindness among diabetic patients, necessitates early detection for effective treatment. This project introduces an explainable and semi-supervised approach to DR classification using a hybrid deep learning framework that leverages the strengths of both convolutional and graph-based neural architectures. Initially, pretrained DesnseNet preprocessing is applied to normalize image inputs, enhancing consistency across diverse retinal fundus images. Feature extraction is then performed using MobileNet, a lightweight and efficient convolutional neural network optimized for real-time medical applications. To further capture the structural relationships between image features, a Graph Convolutional Network (GCN) is integrated with MobileNet in a unified model. The final classification spans five DR stages: No_DR, Mild, Moderate, Severe, and Proliferative_DR. The model is trained and validated on the Kaggle APTOS dataset and deployed through a Flask-based web interface, enabling users to upload retinal images and receive immediate diagnostic predictions. The proposed architecture enhances diagnostic accuracy and interpretability while maintaining computational efficiency, making it viable for clinical settings and remote screenings. This system demonstrates significant potential in aiding early DR diagnosis and improving patient care outcomes.   Keywords: Diabetic Retinopathy, MobileNet, DenseNet, GCN, Hybrid Model, Deep Learning, Retina Screening, Flask, Kaggle Dataset, Medical Imaging, Explainable AI.