An Efficient Deep Learning Approach for Diabetic Retinopathy Detection from Retinal Fundus Images using CNN, LSTM, and Compressive Sensing Techniques

Abstract

Early detection and precise classification are vital for early intervention DR, one of the most common consequences of diabetes mellitus that can affect the eyesight. This research aims to create and evaluate a novel deep learning system for automated DR identification using retinal fundus images. The system will be built on (CS), (CNNs), and (LSTM) networks. The suggested CS-CNN-LSTM model integrates compressive sensing, (CNNs), and (LSTMs) to reduce the amount of picture data, extract spatial features, and learn sequential features. The model's performance is evaluated using metrics such as accuracy, precision, sensitivity, specificity, F1-score, and compression efficiency. The framework encompasses five disease categories: No DR, Mild DR, Moderate DR, Severe DR, and Proliferative DR. Accuracy (96.20%), precision (95.60%), sensitivity (96.40%), specificity (97.10%), and F1-score (96.10%), according to the presented findings, are all achieved by the suggested model. When the compression level is 50%, the storage reduction is 50.00% and the classification accuracy is 94.80%, while for 75% compression, the storage reduction is 75.00% and the classification accuracy is 91.60%. Based on comparative evaluation, it is found that the proposed model shows better overall classification performance than the evaluated SVM, conventional CNN, CNN-LSTM and CS-CNN model. The results indicate that the combination of compressive sensing with CNN and LSTM networks is a promising method to achieve a trade-off between diagnostic accuracy and data efficiency for automated DR screening. The model needs to be validated with independent sets of data and reproducible experimental testing to ensure its reliability and clinical applicability.