Abstract:Diabetic retinopathy (DR) is the most common microvascular complication of diabetes and a leading cause of vision loss among the working-age population. Conventional screening methods, which depend on manual interpretation of fundus images by ophthalmologists, face limitations including uneven distribution of medical resources and subjectivity in diagnosis. In recent years, deep learning (DL) technology, with its strengths in image recognition and feature learning, has emerged as a novel and efficient automated approach for early DR screening. This article provides a systematic review of DL applications in DR screening, covering technical principles, mainstream algorithms, dataset construction, model training strategies, and the potential for integrating multi-modal data. It further examines key challenges in current applications, such as data quality, model interpretability, system integration, and obstacles to clinical translation. Finally, future directions are discussed, including the development of lightweight models, multi-disease joint prediction, and the establishment of interdisciplinary collaborative frameworks, with the aim of supporting the clinical adoption of DL-based DR screening.
