Diagnostic accuracy of AI or DL-enhanced technologies in the diagnosis of diabetic retinopathy: a systematic review

Abstract Background Diabetic retinopathy (DR) remains a leading cause of preventable vision loss globally, particularly among individuals with long-standing diabetes. Conventional DR screening methods, such as manual retinal grading by ophthalmologists, are labor-intensive, time-consuming, and subject to inter-grader variability. These limitations highlight the urgent need for automated, scalable, and reliable diagnostic solutions that can improve early detection and the need for clinical detection. Objective This systematic review aimed to evaluate the diagnostic accuracy of artificial intelligence (AI) and deep learning (DL) models for DR detection, with a particular focus on hybrid architectures that combine convolutional neural networks (CNNs), Vision Transformers (ViTs), and traditional machine learning (ML) classifiers. Methods A comprehensive search of six databases covering 2012–2024 identified 327 studies, of which 20 met the inclusion criteria. Two independent reviewers conducted screening, data extraction, and risk of bias assessment using the QUADAS-2 tool. Due to heterogeneity in model architectures, datasets, and reporting, a meta-analysis was not feasible. Complete database search strategies are available in the supplementary appendix. Results Hybrid AI models consistently outperformed conventional DL and ML approaches, achieving sensitivities of 92–100% and specificities above 80%. Models that integrated both fundus photography and optical coherence tomography (OCT) images achieved the highest diagnostic accuracy (sensitivity 97.2%). Attention mechanisms, ensemble learning, and explainable AI (XAI) methods (e.g., Grad-CAM, LIME) enhanced model interpretability and clinical applicability. Conclusions Hybrid AI-based diagnostic platforms show strong potential to transform DR screening by offering improved scalability, accuracy, and cost-effectiveness. However, clinical translation will require greater data diversity, robust external validation, and regulatory compliance. These findings provide timely evidence to support the integration of AI-driven DR screening into national diabetes strategies and public health policies, particularly in regions with high diabetes prevalence, such as Qatar and the Middle East.