ObjectiveTo systematically review the value of deep learning (DL) on the diagnosis of diabetic retinopathy (DR) based on color fundus photographs. MethodsThe PubMed, Embase, Web of Science, Cochrane Library, IEEE, CNKI, VIP, WanFang Data databases were systematically searched to collect the studies on the use of DL in the diagnosis of DR from January 2019 to November 2024. Two reviewers independently screened literature, extracted data and assessed the risk of bias of the included studies. Meta-analysis was then performed by using RevMan 5.4.1, Meta-Disc 1.4 and Stata 16.0 software. ResultsA total of 16 literatures were included, involving 215 560 images. Meta-analysis results showed that the combined sensitivity of DL in diagnosing DR was 0.97 (95%CI 0.94 to 0.98), the specificity was 0.97 (95%CI 0.94 to 0.98), the AUC was 0.99 (95%CI 0.94 to 0.98), and the DOR was 852 (95%CI 403 to 1 803). ConclusionDL has a high diagnostic value for DR. However, there is a high degree of heterogeneity among different studies. In the future, more large-sample, high-quality studies can be included to confirm its clinical applicability.
