Diagnostic performance of deep learning models in retinal vein occlusion: a systematic review and meta-analysis.

BACKGROUND

The diagnosis of retinal vein occlusion (RVO) requires specialised imaging and clinician expertise. This systematic review and meta-analysis aimed to quantitatively synthesise the current state of artificial intelligence models dedicated to RVO diagnosis.

METHODS

Cochrane Central, Embase, Ovid Medline, and PubMed were searched from inception to June 28, 2025. We included studies that evaluated deep learning models for RVO diagnosis and were validated on clinical datasets. The primary outcomes were the accuracy, sensitivity, and specificity of deep learning models for RVO diagnosis. Risk of bias was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 tool.

RESULTS

Of 612 studies screened, 27 met the inclusion criteria, and 15 were included in the meta-analysis. Pooled deep learning model performance for RVO diagnosis was as follows: 97.4% accuracy (95% confidence interval [CI] 95.0%-98.7%), 89.9% sensitivity (95% CI 85.3%-93.3%), and 98.6% specificity (95% CI 96.5%-99.4%). Models predominantly used colour fundus photography (n = 18), OCT (n = 6), and ultra-widefield imaging (n = 2). Despite substantial heterogeneity, overall study quality was acceptable. In the Grading of Recommendations Assessment, Development, and Evaluation assessment, reasons for downgrading were inconsistency and imprecision.

CONCLUSIONS

Deep learning models showed high diagnostic accuracy in detecting retinal vein occlusion across different imaging modalities. However, the certainty of the evidence, assessed using Grading of Recommendations Assessment, Development, and Evaluation, was low for all outcomes. Future research should investigate the feasibility and costs of implementing deep learning for RVO diagnosis in real-world clinical environments.