Using AI to Identify Diabetic Retinopathy and Clinically Significant DME


Artificial intelligence could be the future of detecting certain diseases in the eye, after a study presented at ASRS 2019 found that the EyeArt system was effective in detecting referable diabetic retinopathy from images.

Jennifer Lim, MD

Jennifer Lim, MD

Results of a prospective study demonstrating the sensitivity and specificity of the EyeArt system to detect moderate nonproliferative or clinically significant diabetic macular edema among patients were presented at the 2019 American Society of Retina Specialists Annual Meeting. 

Results of the study, which found that the sensitivity and specificity of the EyeArt system were high enough to detected referable diabetic retinopathy in patients, were presented on Tuesday by Jennifer Lim, MD, professor of ophthalmology and director of the Retina Service at University of Illinois Eye and Ear Infirmary. 

To determine whether an artificial intelligence (AI) system would have high enough sensitivity and specificity to identify diabetic retinopathy, investigators conducted a prospective, multicenter study that enrolled patients from 15 centers. Patients enrolled in the study underwent undiluted 2-field, 45 degree, funds photography and then dilated 4-wide field stereoscopic funds photography.

The EyeArt system provided investigators eye level results about referable diabetic retinopathy, which was defined as moderate nonproliferative diabetic retinopathy or higher per International Clinical Diabetic Retinopathy severity scale or as clinically significant diabetic macular edema.

For instances when the EyeArt system was unable to grade the images, the 2-field photos were repeated after dilation. Investigators noted that dilated wide field photographs were the reference standard and were graded by Wisconsin Fundus Photograph Reading Center graders using the Early Treatment Diabetic Retinopathy Study severity scale. Investigators compared detection rates of the AI system compared to the reference standard.

The study consisted of 1822 eyes from 911 patients. A cohort of 1674 patients had both gradable 2-field and 4-field images. Of those 1674 eyes, 310 were positive and 1364 were negative for referable diabetic retinopathy. Using only undiluted images, the sensitivity of the EyeArt system was 95.5% (95% CI; 92.4% to 98.5%), the sensitivity was noted as 86% (95% CI: 83.7% to 88.4%), and the gradeability rate was 87.5% (95% CI; 85.4% to 89.7%).

Additionally, Investigators noted that dilated 2-field photos were required for 214 eyes — of those, 170 were considered gradable and 44 remained ungradable on the EyeArt system. Using the established dilated-if-ungradable protocol, the EyeArt system’s gradeability increased to 97.4% (95% CI; 96.4% to 98.5%), sensitivity increased to 95.5% (95% CI; 92.6% to 98.4%), and specificity increased to 86.5% (95% CI; 84.3% to 88.7%).

Investigators noted 14 false negative referable diabetic retinopathy eyes, 14 of which had moderate ETDRS level 35 and 1 had clinically significant diabetic macular edema. Additionally, investigators noted 184 instances of false positive referable diabetic retinopathy, 119 of which had mild diabetic retinopathy and 20 had non-diabetic retinopathy conditions, which included age-related macular degeneration, vein occlusion, epiretinal membrane, vitreous opacities, optic disc edema, atrophy, scar, or nevus. 

This study, titled “Artificial Intelligence Screening for Diabetic Retinopathy: Analysis from a Pivotal Multi-center Prospective Clinical Trial,” was presented at ASRS 2019.

Related Videos
Hong-Uyen Hua, MD; Sruthi Arepalli, MD; and Peter Kaiser, MD
Connective Tissue Disease Brings Dermatology & Rheumatology Together
Kelly Nichols, OD; Laura Periman, MD; and Mile Brujic, OD
What Makes JAK Inhibitors Safe in Dermatology
© 2024 MJH Life Sciences

All rights reserved.