A Comprehensive Review of Convolutional Neural Network-Based Methods for Diabetic Retinopathy Detection and Classification in Fundus Images

Diabetic Retinopathy (DR) is a leading cause of preventable blindness worldwide. Early detection through fundus image screening is essential; however, manual grading by ophthalmologists is time-consuming, subjective, and limited by specialist shortages. In recent years, Convolutional Neural Networks (CNNs) have demonstrated expert-level performance in automated DR detection and classification from fundus images. This review comprehensively examines CNN-based methods for DR detection and classification. We analyze widely used public datasets (e.g., APTOS ۲۰۱۹, Eye PACS, Messidor), common preprocessing techniques, and state-of-the-art CNN architectures including VGG, ResNet, Inception, DenseNet, and EfficientNet. Special emphasis is placed on transfer learning and ensemble approaches, which have significantly enhanced classification accuracy. Regulator-approved systems have reported pooled sensitivity of ۹۳% and specificity of ۹۰% for referable DR. Nevertheless, key challenges remain, including class imbalance, poor model generalization across populations and devices, lack of interpretability (the "black box" problem), and difficulties in clinical deployment. Future directions include the development of lightweight models for mobile screening, explainable AI (XAI), multimodal imaging, federated learning, and creation of more diverse datasets. CNN-based systems show strong potential to make DR screening more accessible, efficient, and equitable globally.

کلیدواژه ها:

Diabetic Retinopathy ، Convolutional Neural Networks ، Fundus Images ، Transfer Learning ، Deep Learning ، Medical Image Analysis ، Computer-Aided Diagnosis