Performance of Deep Learning in Classifying Age-Related Macular Degeneration From Images: Systematic Review and Meta-Analysis
Source: PubMed Central Open Access, NCBI / U.S. National Library of Medicine
Abstract Background Age-related macular degeneration (AMD) is a leading cause of irreversible blindness worldwide. Retinal imaging and deep learning (DL) may support scalable screening, but deployment requires evidence on pooled performance. This is important because missed neovascular disease may delay treatment, whereas excessive false positives may overload referral pathways. Objective This study aimed to compare the diagnostic performance of DL algorithms with ophthalmologists for detecting AMD and differentiating wet AMD (wAMD) from dry AMD (dAMD) and to identify factors that influence DL performance. Methods PubMed, Embase, Web of Science, and the Cochrane Library were searched through October 5, 2025, and updated on April 19, 2026. Eligible studies applied DL to classify AMD from normal retinas or wAMD from dAMD using retinal images. Two reviewers (MHT and XL) independently extracted data and assessed risk of bias using the Prediction model Risk Of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI) tool. Pooled sensitivity, specificity, accuracy, and area under the curve were estimated using bivariate random-effects models. Clinician comparisons were stratified by experience (junior vs senior). Small-study effects were assessed via Deeks’ funnel plot asymmetry test. Evidence certainty was appraised using the Grading of Recommendations, Assessment, Development, and Evaluation framework. The protocol was registered in the International Prospective Register of
Abstract
Abstract Background Age-related macular degeneration (AMD) is a leading cause of irreversible blindness worldwide. Retinal imaging and deep learning (DL) may support scalable screening, but deployment requires evidence on pooled performance. This is important because missed neovascular disease may delay treatment, whereas excessive false positives may overload referral pathways. Objective This study aimed to compare the diagnostic performance of DL algorithms with ophthalmologists for detecting AMD and differentiating wet AMD (wAMD) from dry AMD (dAMD) and to identify factors that influence DL performance. Methods PubMed, Embase, Web of Science, and the Cochrane Library were searched through October 5, 2025, and updated on April 19, 2026. Eligible studies applied DL to classify AMD from normal retinas or wAMD from dAMD using retinal images. Two reviewers (MHT and XL) independently extracted data and assessed risk of bias using the Prediction model Risk Of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI) tool. Pooled sensitivity, specificity, accuracy, and area under the curve were estimated using bivariate random-effects models. Clinician comparisons were stratified by experience (junior vs senior). Small-study effects were assessed via Deeks’ funnel plot asymmetry test. Evidence certainty was appraised using the Grading of Recommendations, Assessment, Development, and Evaluation framework. The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251243276). Results Overall, 28 studies were included, comprising 77,485 samples for AMD detection and 28,705 samples for wAMD versus dAMD classification. For AMD detection, DL achieved a pooled sensitivity of 0.98 (95% CI 0.96‐0.99; prediction interval [PI] 0.95‐0.99), specificity of 0.98 (95% CI 0.95‐0.99; PI 0.95‐0.99), accuracy of 0.97 (95% CI 0.96‐0.99), and area under the curve of 1.00 (95% CI 0.99‐1.00). For wAMD versus dAMD, DL showed sensitivity of 0.95 (95% CI 0.91‐0.97; PI 0.89‐0.97), specificity of 0.95 (95% CI 0.93‐0.97; PI 0.92‐0.97), accuracy of 0.95 (95% CI 0.92‐0.97), and area under the curve of 0.99 (95% CI 0.97‐0.99). DL showed higher sensitivity than senior ophthalmologists for AMD (0.98 vs 0.75;<.001) and higher specificity and accuracy than junior ophthalmologists for wAMD classification. Optical coherence tomography–based models performed more consistently than color fundus photography or multimodal models. Evidence certainty was moderate. Conclusions Compared with ophthalmologists, DL algorithms demonstrated superior and more balanced diagnostic performance in the available head-to-head evidence, potentially providing a consistent decision-support baseline that mitigates human threshold-dependent trade-offs. However, high heterogeneity, wide PIs, predominantly retrospective designs, and possible performance inflation from internal validation mean that these relative performance findings remain preliminary rather than deployment ready. DL should be viewed as a triage adjunct requiring local calibration, not an autonomous diagnostic replacement. Prospective, multicenter, patient-level external validation with prespecified human comparison arms is required.