Smart Diagnosis of Diabetic Retinopathy Using Deep Learning and Explainable AI

The “Smart Diagnosis of Diabetic Retinopathy” project presents a deep learning–based automated system for the early detection and classification of diabetic retinopathy (DR), a leading cause of blindness in diabetic patients. The system utilizes EfficientNet-B5, a high-performing and computationally efficient convolutional neural network (CNN), to classify retinal fundus images into five clinically relevant stages: No DR, Mild, Moderate, Severe, and Proliferative DR. The model is trained on the APTOS 2019 Blindness Detection dataset consisting of over 3,500 annotated images and achieves a best validation accuracy of 72.17%. To enhance image quality and model performance, preprocessing techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and Retinex-based enhancement are applied to improve contrast and feature visibility in retinal scans. A key component of this system is explainable AI integration, specifically Gradient-weighted Class Activation Mapping (Grad-CAM), which generates intuitive heatmaps highlighting regions of interest that influenced the model’s prediction. This improves interpretability and supports clinical decision-making by offering visual transparency into the AI’s diagnostic reasoning. The model outputs not only the predicted DR stage but also the associated confidence score. Additionally, the application includes heuristic-based validation to filter out non-fundus inputs, ensuring that only relevant medical images are processed. The complete system is deployed using a Gradio-based interactive web interface hosted on Hugging Face Spaces, enabling real-time image upload, classification, and visualization. Designed for scalability and integration into telemedicine platforms, the solution is particularly suited for resource-limited settings, offering a cost-effective, non-invasive, and interpretable tool for early DR screening and long-term diabetic eye care management.