AI and big data have become enabling technologies for diabetic retinopathy (DR) screening.
Diabetic retinopathy is the most common microvascular complication of diabetes and the leading cause of blindness among the working‑age population worldwide.With population aging and the rising prevalence of diabetes, the situation for DR prevention and treatment has become increasingly severe.
AI and big data are key enablers for early DR screening, and a complete system has been formed covering technical validation, clinical application, and standardization.The Automated Retinal Imaging Analysis System (ARIAS) serves as the core tool for DR screening, and multiple large‑sample studies have confirmed its high diagnostic performance.A real‑world study in England involving more than 200,000 participants showed that several commercial ARIAS systems achieved a sensitivity of over 95% for detecting moderate‑to‑severe and proliferative DR, regardless of age or ethnicity, supporting their use as a first‑line triage tool.A recent meta‑analysis demonstrated that deep learning‑based ARIAS achieved high performance in detecting vision‑threatening DR (AUC = 0.974) and is cost‑effective in high‑income countries.
Multimodal ocular imaging AI models have achieved clinical translation.EyeFM, a newly developed multimodal foundation model for ocular imaging in China, integrates five core imaging modalities, including color fundus photography, OCT, and ultra‑widefield fundus imaging.Through vision‑language pre‑training, it can accurately identify lesions, generate standardized reports, and respond to medical queries.A global multicenter randomized controlled trial (RCT) verified that it significantly improves diagnostic accuracy in primary care settings.The EyeCLIP multimodal vision foundation model from The Hong Kong Polytechnic University addresses the challenge of rare lesion recognition through multimodal contrastive learning.
Furthermore, AI can analyze retinal thickness maps to clarify correlations with genomic variations, metabolites, and systemic diseases,making the retina a window for assessing systemic metabolic and immune status and expanding the value of DR screening.
Standardization and ethical guidelines are core guarantees for the large‑scale deployment of AI‑based DR early‑screening technologies.The lack of DICOM standards for retinal imaging has limited cross‑center development of AI models,while the launch of the AI‑READI standardized dataset provides unified support for multimodal DR research.Issues such as the AI “black box” effect, data privacy, and algorithmic bias require collaborative governance among regulators, healthcare professionals, and developers to establish a normative framework and ensure the safe implementation of these technologies.