Comparative Analysis of AI and Ophthalmologist Grading in Diabetic Retinopathy Detection

Background: Diabetic retinopathy (DR) poses a significant global health challenge that needs scalable and efficient screening pathways beyond the current limitations of teleophthalmology. This study retrospectively evaluated the diagnostic performance of an artificial intelligence (AI) DRISTi system (Version 2.1) against ophthalmologist grading for more-than-mild diabetic retinopathy (mtmDR), vision-threatening diabetic retinopathy (vtDR), and diabetic macular edema (DME). Methods: The methods involved a retrospective, observational, non-interventional validation comparing the AI DRISTi system's output to ophthalmologist grading on 739 colour fundus images acquired using Topcon NWC 400, CrystalVue NFC 600/700, Canon CR2/CR2 AF, and Zeiss VISUCAM 500 cameras. Results: Primary outcomes included sensitivity and specificity, with statistical analyses utilizing 2 × 2 contingency tables and 95% confidence intervals. The AI system achieved an accuracy of 93.36% (sensitivity 95.03%; specificity 92.90%) for mtmDR, 98.64% (sensitivity 96.92%; specificity 99.01%) for vtDR, and 97.97% (sensitivity 92.85%; specificity 98.88%) for DME. Performance was robust and consistent across all evaluated camera types. Conclusions: In conclusion, the AI DRISTi system (Version 2.1) demonstrates strong diagnostic performance for mtmDR, vtDR, and DME, comparable to leading commercial AI systems, from fundus photographs acquired across multiple camera platforms. This system holds significant promise as an adjunctive screening tool for large-scale DR screening programs, contributing to early detection, appropriate triage, and the prevention of vision loss in at-risk populations.