ML and DL for Diabetic Retinopathy Detection: A Systematic Review

Background: Diabetic retinopathy (DR) is a major cause of preventable blindness worldwide. Early detection and timely treatment are vital to avoid irreversible vision loss. Artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), has transformed automated DR screening, providing scalable, accurate, and cost-effective solutions. Objective: This systematic review synthesizes advances in ML and DL approaches for DR detection, critically evaluating datasets, methods, metrics, explainability, and translational barriers. Methods: Following PRISMA guidelines, a systematic search was conducted in PubMed, IEEE Xplore, Scopus, and Web of Science between January ۲۰۱۶ and March ۲۰۲۵. Studies applying ML/DL to DR detection or grading using fundus or OCT images were included. Of ۱,۱۲۴ records identified, ۹۲ articles met eligibility criteria. Results: Classical ML methods such as support vector machines, random forests, and k-means clustering achieved moderate results with handcrafted features but lacked scalability. DL, particularly convolutional neural networks (CNNs) like InceptionNet, ResNet, DenseNet, and EfficientNet, consistently achieved sensitivity above ۹۰% on large datasets. Vision Transformers (ViTs) and CNN–ViT hybrids improved global feature learning and interpretability. Ensemble frameworks reached state-of-the-art accuracy, while lightweight mobile models showed potential for low-resource deployment. Explainability methods such as Grad-CAM and saliency maps enhanced clinical trust, though interpretability remains difficult. Conclusions: DL dominates current DR detection research, surpassing ML. Remaining challenges include dataset bias, reproducibility, and real-world validation. Future work should prioritize federated learning, multimodal integration, domain adaptation, and regulatory-approved clinical translation.

کلیدواژه ها:

Diabetic retinopathy ، deep learning ، machine learning ، medical image analysis