Deep Learning-Based Framework for Diabetic Disease Progression Prediction Using Retinal Fundus Images

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness worldwide, making its early detection and accurate classification critical for effective treatment and prevention. However, traditional diagnostic methods are time-consuming and reliant on subjective clinical expertise, leading to inconsistent outcomes. We present OptiRetina-Net, an effective deep learning model, to tackle this difficulty. It uses long-short term memory networks for temporal analysis and convolutional neural networks for spatial feature extraction. This type of architecture is useful in capturing fine details of the retinal structures and temporal variations in the disease state for DR staging. Using a 70:15:15 split between training, validation and testing, the research used a balanced dataset of 10,000 labelled retinal pictures classified by DR severity (No DR, Mild, Moderate, Severe, Proliferative DR). The use of Recursive Feature Elimination (RFE) and feature importance obtained from SHAP analysis allowed for focusing on the clinically meaningful predictors only. Grid search was used for hyperparameter tuning and early stopping was employed to avoid overtraining, while k-fold cross validation applied for validation. For the testing set, OptiRetina-Net yielded an overall accuracy of 88 percent and an AUC of 0.91, with 95 percent accuracy on No DR and 82 percent on Proliferative DR. To support this, interpretability tools like Grad-CAM and SHAP offered visual and numerical information about the model’s decision-making process, in line with clinical significance. The findings prove that the proposed framework can be used for early identification of DR and monitoring of its progression. The proposed system has a future scope of applications in telemedicine and real-time clinical decision support systems and it also provides a clear explanation of the features of the diagnosed DR.