AI system detects referable diabetic retinopathy at higher rate than retina specialists
A study comparing general ophthalmologists, retina specialists, and an AI screening system (EyeArt) for detection of referable diabetic retinopathy using fundus photo standards for comparison found that the AI system had the highest sensitivity for detection, according to a presentation at the 2020 ASRS Virtual Meeting.
With the number of patients with diabetes growing steadily, ophthalmologists face a huge screening burden, said Jennifer I. Lim, MD during the presentation. Adoption of an AI screening system will be both cost-effective and timely.
A total of 521 patients underwent dilated ophthalmoscopy by non-retina specialists (n = 406) or retina specialists (n = 115). Fundus Photograph Reading Center (FPRC) gradings found 207 positive eyes including 190 moderate NPDR, 1 severe NPDR, 15 PDR, and 37 CSDME.
For undilated images, AI sensitivity, specificity, and gradeability rate was 96.1%, 87.5%, and 85.3%, respecitvley.
For the dilated images required for 147 eyes, AI gradeability rate improved to 97.4%, sensitivity to 96.4%, and specificity to 88.4%. A total of 26 eyes remained ungradable by the system.
Ophthalmoscopy sensitivity and specificity was 27.7% and 99.6%, respectively; retina specialists sensitivity and specificity was 59.5% and 98.9%, respectively; and non-retina specialists sensitivity and specificity was 20.7% and 99.8%, respectively.
“When we look at the retina specialist compared to the general ophthalmologists and the EyeArt system both the retina specialists and the EyeArt system did not miss vision threatening retinopathy, that is severe NPDR or higher grades of retinopathy,” noted Dr Lim.
Lim J, et al. Artificial intelligence screening for diabetic retinopathy: subgroup comparison of the EyeArt system with ophthalmologists’ dilated exams. Presented at: 2020 ASRS Virtual Meeting.
Grandin Library Building
Six Leigh Street
Clinton, New Jersey 08809