Evaluation of the implementation of an approved artificial intelligence system for the detection of diabetic retinopathy

Introduction The aim of this study was to evaluate the accuracy of an artificial intelligence (AI)-based analysis of fundus photographs compared to ophthalmologists in diabetic retinopathy screening in an internal medicine clinic. In addition, the total examination time as well as the patient and examiner satisfaction were surveyed. Methods In the study, 112 outpatients received fundus photography with automated diagnosis of diabetic retinopathy (DR) via the IDx-DR system (Digital Diagnostics). The images were taken with the Topcon TRC-NW400 camera (Topcon Corp. Japan). Inclusion criterion was a diagnosis of diabetes mellitus type 1, 2, or 3. Patients who could not be imaged with sufficient quality in miosis were imaged in mydriasis. Results Of 112 patients, analysis was possible in 107 patients (95.5 %) by grading by IDx-DR, based on fundus images - vs. 103 patients (91.9 %) by grading by ophthalmologists, based on the same, high-resolution fundus images. In the remaining patients, grading was possible by funduscopy alone. There was a highly significant correlation regarding the assessment of the severity of diabetic retinopathy between the examiner and the IDx-DR system (Correlation coefficient (r) = 0.8738; p < 0.0001). Patient satisfaction was 4.5 +/- 0.6 [1-5], and total examination time in miosis averaged 3:04 +/- 0: 28 [min:sec]. Conclusion Retinopathy screening using IDx-DR enables automated diagnosis based on fundus photographs with a robust, technical and clinical workflow that allows timely and reliable assessment of diabetic retinopathy and is associated with high patient satisfaction.