Diagnostic Accuracy Of Artificial Intelligence-Based Fundus Imaging For Detecting Diabetic Retinopathy: A Systematic Review And Meta-Analysis

Background: The problem of diabetic retinopathy (dr) is among the significant reasons causing blindness that can be avoided in the global context, and screening is a significant aspect that helps to identify this situation at the initial stage. Conventional screening is very reliant on the ophthalmologists which creates workforce bottlenecks in the high-prevalence regions. A new promising solution has emerged in the fundus imaging with artificial intelligence (ai) but the diagnostic accuracy of the tool is different in various studies. In this meta-analysis and systematic review, it was desired to evaluate the diagnostic properties of ai models in the identification of dr. Methods: All the studies concerning the topic and published in january 2018 to september 2025 were found under the following guidelines of prisma 2020 using a systematic search in pubmed, embase, web of science, scopus, ieee xplore, and cochrane library. The inclusion criteria were adult patients who receive fundal imaging and the ai systems were contrasted with the results of ophthalmologists or etdrs reference. The primary outcomes were sensitivity, specificity and area under the curve (auc); the secondary ones were diagnostic odds ratios (dor), age, imaging modality, and ai model type subgroups. Quality was assessed using cochrane rob 2.0 rct evaluation tool and newcastle-ottawa scale to assess quality based on observational studies. The pooled estimates were obtained through random-effects meta-analysis. Results: Eighteen articles were incorporated and comprised 1782 patients/eyes. Pooled estimates of dor, auc and sensitivity respectively were 0.92 (95% ci: 0.89-0.95), 0.87 (95% ci: 0.83-0.89), and 0.96 (95% ci: 0.94-0.98) for specificity. Subgroup analysis was more precise in patients that are older (>65 years, sensitivity 93.5%), and wide-field fundus imaging (auc 0.96). The imaging of smartphone was of a small and yet clinically plausible accuracy (auc 0.91). The heterogeneity level was medium (i 2 = 42- 55 percent), and sensitivity analysis did not disprove the power of pooled estimates. Risk of bias was predominantly low-to-moderate and it did not show much evidence of publication bias. Conclusion: Within fundus imaging, ai-based systems are very accurate in diagnostic identification of dr, possess good pooled performance measures and are resistant to patient subgroups or imaging modalities. Their scalability and low discontinuation considerations render viability to real-life screening particularly where the burden is enormous or there is a resource limitation. Future studies need to be focused on longitudinal results, smartphone imaging and integration into multimodal care pathways to support evidence of high scale adoption.

Перевод пока недоступен