Deep Learning for Early Detection and Classification of Diabetic Retinopathy in Clinical Practice

Abstract

Early detection of diabetic retinopathy (DR) is crucial to prevent vision loss. In recent years, deep learning has revolutionized automated DR screening by learning complex retinal patterns. Convolutional neural networks (CNNs) are commonly used to extract spatial features from fundus images, while some advanced models also incorporate sequential analysis (e.g. RNNs) to track disease progression over time. In this paper, we review and develop deep learning frameworks for DR detection, focusing on CNN and hybrid CNN-RNN models. We conducted experiments on publicly available datasets (EyePACS/Kaggle, APTOS, Messidor, DRIVE) to evaluate model performance. Our deep CNN and transfer-learning models achieve high sensitivity and specificity. For example, a hybrid CNN-LSTM model (TAHDL) attains ~97-99% accuracy across datasets. We also propose a lightweight CNN (RSG-Net) that classifies DR into multiple grades with 99.4% accuracy on Messidor. Key factors include preprocessing (contrast enhancement, augmentation) and addressing data imbalance. These models can assist clinicians by providing fast, reliable DR screening.