Diagnostic Accuracy of Deep Learning Models for Early Diabetic Retinopathy Detection from Fundus Images: A Systematic Review and Meta-Analysis

Abstract

Background: Diabetic retinopathy (DR) remains a major cause of preventable visual loss, while access to trained retinal graders and ophthalmologists is uneven. Deep learning systems can classify color fundus photographs at the point of screening and may enable earlier referral before progression to vision-threatening disease.
Objective: To synthesize high-quality external and clinical validation evidence for deep learning detection of actionable DR and to estimate pooled sensitivity and specificity.
Methods: A focused systematic review was conducted for peer-reviewed external, prospective, pivotal and post-deployment validations published through December 2025. The quantitative synthesis included eight independent validation cohorts from six Q1-journal reports in which sensitivity, specificity and 95% confidence intervals were available for referable, moderate-or-worse or more-than-mild DR. Sensitivity and specificity were logit transformed and pooled separately using DerSimonian-Laird random-effects models. Heterogeneity was summarized with I2. 
Results: Across the eight cohorts, sensitivity ranged from 87.0% to 95.5% and specificity from 85.0% to 98.5%. Pooled sensitivity was 90.5% (95% CI, 88.8%-91.9%; I2=52.5%), while pooled specificity was 93.9% (95% CI, 90.8%-96.1%; I2=98.2%). Leave-one-out analyses produced pooled sensitivity estimates of 90.1%-90.8% and specificity estimates of 92.7%-94.7%, indicating stable average performance but substantial threshold-related heterogeneity in specificity. Later real-world and post-deployment studies continued to report sensitivities above 90% for sight-threatening or severe DR, although ungradable images, disease threshold, camera type and workflow design materially affected performance.
Conclusion: Deep learning analysis of fundus photographs provides high sensitivity and generally high specificity for identifying actionable DR, supporting its use as a screening and referral technology. The principal implementation challenge is no longer whether these systems can detect disease, but how programs manage operating thresholds, image gradability, population shift and false-positive referral burden.