Evaluation of AI-enhanced non-mydriatic fundus photography for diabetic retinopathy screening

Abstract

Objective

To assess the feasibility of using non-mydriatic fundus photography in conjunction with an artificial intelligence (AI) reading platform for large-scale screening of diabetic retinopathy (DR).

Methods

In this study, we selected 120 patients with diabetes hospitalized in our institution from December 2019 to April 2021. Retinal imaging of 240 eyes was obtained using non-mydriatic fundus photography. The fundus images of these patients were divided into two groups based on different interpretation methods. In Experiment Group 1, the images were analyzed and graded for DR diagnosis using an AI reading platform. In Experiment Group 2, the images were analyzed and graded for DR diagnosis by an associate chief physician in ophthalmology, specializing in fundus diseases. Concurrently, all patients underwent the gold standard for DR diagnosis and grading—fundus fluorescein angiography (FFA)—with the outcomes serving as the Control Group. The diagnostic value of the two methods was assessed by comparing the results of Experiment Groups 1 and 2 with those of the Control Group.

Results

Keeping the control group (FFA results) as the gold standard, no significant differences were observed between the two experimental groups regarding diagnostic sensitivity, specificity, false positive rate, false negative rate, positive predictive value, negative predictive value, Youden's index, Kappa value, and diagnostic accuracy (X² = 0.371, P > 0.05).

Conclusion

Compared with the manual reading group, the AI reading group revealed no significant differences across all diagnostic indicators, exhibiting high sensitivity and specificity, as well as a relatively high positive predictive value. Additionally, it demonstrated a high level of diagnostic consistency with the gold standard. This technology holds potential for suitability in large-scale screening of DR.