Vitreoretinal disease detection using artificial intelligence: a systematic review and meta-analysis

Early detection of vitreoretinal diseases (VRDs) is critical for preventing vision loss, and currently relies on the examination and interpretation of multimodal imaging techniques. Artificial intelligence (AI) is emerging as a powerful tool to detect abnormalities in vitreoretinal morphology and ideally detect changes at earlier stages to allow for intervention. This meta-analysis evaluates and summarizes the diagnostic performance of AI models in the detection of VRDs using retinal imaging systems.

This study was registered in PROSPERO (CRD42023450207). A comprehensive electronic search of PubMed/MEDLINE, EMBASE, and Web of Science was conducted by three independent reviewers up to August 2023. Study validity was assessed using the QUADAS-2 tool, which evaluates risk of bias across four domains and applicability concerns across three domains. Eligible articles were categorized into nine VRD subgroups-age related macular degeneration, diabetic retinopathy, retinal vascular diseases, retinal dystrophies, Cystoid macular edema, vitreoretinal interface disorders, retinal detachment, Central serous chorioretinopathy, and myopic retinopathy-and included in the meta-analysis. Pooled estimates of accuracy (PEA), sensitivity (PESen), and specificity (PESpe) were calculated for all selected studies.

A total of 195 studies were included in the final analysis, yielding an overall PEA of 95.76% (95% CI: 95.0-96.47), PESen of 91.94% (95% CI: 90.72-93.08) and PESpe of 96.09% (95% CI: 95.27-96.79). In the subgroup analysis, most AI models had a PEA > 90%, especially convolutional neural networks (CNN), followed by support vector machine (SVM) and random forest (RF).

AI diagnostic tools, particularly CNNs, have demonstrated robust performance in VRDs detection. However, results from studies with limited generalizability should be applied cautiously in real-world settings. Further exploration of emerging models, such as large language models (LLMs), is recommended.