Abstract
Purpose :
This study aims to investigate the efficacy of artificial intelligence (AI) grading for determining the severity of diabetic retinopathy (DR) and the severity of diabetic macular edema (DME) on evaluation of smartphone-based retinal imaging.
Methods :
Non-ophthalmic personnel used a smartphone-based retinal camera (RetinaScope) to image the retina of patients with diabetes. The patients were recruited at the University of Michigan (UM) Kellogg Eye Center Retina Clinic and the ophthalmology consultations service at the University of Michigan Hospital after approval from the UM IRB. Artificial intelligence using the Toku Eyes platform analyzed the images and scored DR and markers for DME on scales of 1-5 and 0-4, respectively. Two masked readers independently evaluated the images and scored DR and DME on scales of 1-5 and 1-3, respectively.
Results :
The study included a total of 120 eyes from 70 patients. The mean age of the cohort was 57.0 years (standard deviation 15.7 years); 26 patients (37.1%) were female. The comparison of DR grading between Toku Eyes and the two readers had a kappa of 0.20 (p < 0.0001). Comparison of DME grading had a kappa of -0.0031 (p = 0.6589).
Conclusions :
AI evaluation of DR with Toku Eyes showed slight agreement with independent masked graders. The AI grading of DME showed poor agreement but lacked statistical significance. The level of detail in the severity scales for DR and DME may have contributed to the slight and poor agreement, respectively. Converting scores into referral warranted vs. non-referral warranted categories or clustering scores for comparison may improve results. Alternatively, a larger study may be needed to evaluate the severity scoring more effectively using the current scales.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.