Jennifer Lim, MD: Training & Testing AI to Detect Diabetic Retinopathy

May 1, 2019

After training the EyeArt system on thousands of images of eyes with diabetic retinopathy, investigators tested the AI in a clinical trial.

As the population of patients with diabetes continues to grow, the need for screening for diabetic eye disease increases as well. One potential solution that investigators have turned to is artificial intelligence.

Jennifer I. Lim, MD, Marion H. Schenk Esq. Chair, Professor of Ophthalmology, and Director of Retina Service at the University of Illinois at Chicago and her team have worked to train and test such a system to detect referable diabetic retinopathy in primary care settings.

Lim spoke with MD Magazine® about how the EyeArt system was trained via deep learning and then tested rigorously in a clinical trial.

“In the clinical trial we compared it to the gold standard of these fundus photographs that were read by the Wisconsin Fundus Photograph Reading Center,” said Lim.

In part 1 of the interview, Lim shared about the growing need for diabetic retinopathy screening as the population of patients with diabetes continues to grow worldwide.

How does the EyeArt system detect referable diabetic retinopathy?

How did you design the clinical trial for the EyeArt system?

Well, it uses an artificial intelligence system that was trained using deep neural networks and 375,000 diabetic images. So, basically, you put in: this is a patient whose eye has, say, mild non-proliferative diabetic retinopathy [NPDR], or moderate, or severe—these are different levels that we use—or proliferative diabetic retinopathy and you train the system—this is deep learning. It was then tested on another 175,000 plus images to see how accurate it was and then once that testing was done, we then said, okay, it's ready for use now in a clinical trial.In the clinical trial we compared it to the gold standard of these fundus photographs that were read by the Wisconsin Fundus Photograph Reading Center. So, specifically what we did, is we took 2 pictures without the patient being dilated and these are of a certain size—they're 45-degree images. One is centered on the optic nerve, the other is centered on the macula. So, 2 per eye, and then afterwards we allowed the EyeArt system to interpret the image and say is there is there not this referral level of diabetic retinopathy. So, it had to hit moderate NPDR, specifically, or there had to be swelling with the retina—diabetic macular edema. If for some reason the EyeArt system could not read those images then the patient underwent dilation and these images were repeated, sent again to the EyeArt system to see if it could read it. After that the patient underwent dilation, if they hadn't already been dilated, and that was a vast majority because the vast majority were readable. So, after dilation they then underwent 45-degree wide-field images—4 fields per eye and these 4 images are equivalent to the standard 7-field ETDRS photographs, which have been used in multiple diabetic retinopathy staging studies. We then compared the presence of moderate NPDR or CSDME [clinically significant diabetic macular edema] between the 2 methods—the EyeArt system and the Wisconsin Fundus Photograph Reading Center results.