An Investigation on Optimized Deep Learning Framework for Multi-Stage Diabetic Retinopathy Detection Using Retinal Images

Abstract

Diabetic Retinopathy (DR) remains one of the leading causes of preventable blindness worldwide, particularly among the working-age population. Early and accurate detection of DR is critical for timely clinical intervention and vision preservation. This study proposed an optimized deep learning-based framework for automated detection and multi-stage classification of DR using retinal fundus images. The framework integrated structured preprocessing, data augmentation, and deep feature extraction through Convolutional Neural Networks (CNNs), along with hyperparameter-optimized classification strategies. Publicly available datasets, including MESSIDOR, IDRiD, and ODIR-2019, were utilized to ensure robustness and generalizability. A stratified hold-out partition comprising 60% training, 10% validation, and 30% testing data was adopted to preserve class proportions while retaining a comparatively large independent test set. The proposed system was evaluated using performance metrics such as accuracy, precision, sensitivity, specificity, and Area Under the Curve (AUC). The experimental results demonstrated that the optimized framework achieved a classification accuracy of 98.5%, outperforming conventional CNN (91%) and DCNN (94%) architectures. The ROC analysis yielded an AUC of 0.98, indicating excellent discriminative capability across severity grades. The confusion matrix analysis further confirmed balanced performance with reduced false-positive and false-negative rates. The results suggested that the proposed framework provided reliable, scalable, and computationally efficient DR screening, making it suitable for large-scale automated clinical deployment.