Multi-Modal Detection and Classification of Diabetic Retinopathy

Early detection and precise severity analysis of diabetic retinopathy are crucial to preventing vision loss and improving patient outcomes. To address this challenge, an advanced deep learning model for improved detection and severity analysis of diabetic retinopathy has been proposed, capable of classifying the condition into five severity levels, ensuring higher accuracy and interpretability. To start with, the EfficientNet and GoogLeNet ensemble model was built and tested that reached 95.6% accuracy, which shows the model's stability. Additional improvements were made by adopting attention layers, stacking approaches and HbA1c% levels to increase the accuracy to 98.6%, thus offering a more comprehensive insight into the progression of the disease. The use of explainable AI was performed for interpretability, along with the feature of segmentation of affected retinal regions that helps make accurate medical assessment and decision planning. A user-friendly interface is created to support real time prediction where the clinician can upload the retinal images and get the report back with prediction. This vast framework helps not only to enhance diagnostic efficacy involving diabetic retinopathy, but also equip healthcare practitioners with the crucial instruments for decision-making regarding the results, thus providing a prospect for enhanced control over the disease.