Study of AI for retina disease finds many unusable images

Key Takeaways

• A recent study using artificial intelligence (AI) to detect diabetic retinopathy from retinal photo screenings has found wide disparities in the quality of data being fed into the algorithm. 
• The research suggests that ongoing performance monitoring, and perhaps more automated equipment, may be necessary to enable quality screenings in a primary care setting. 

AI has drawn interest in ophthalmology for its potential to track disease trends in huge populations, such as the 38.4 million people in the United States with diabetes who are at risk for diabetic eye disease. However, a recent study using AI to detect diabetic retinopathy from retinal photo screenings has found wide disparities in the quality of data being fed into the algorithm. Screening photos captured in nine primary care settings were three times more likely to be unusable than those obtained in two ophthalmology clinics (42.5% vs 14.5%, respectively), a study at Temple University in Philadelphia found.  The number of patients diagnosed with more-than-mild diabetic retinopathy also varied significantly between the two settings — 13% in primary care and 24% in ophthalmology — as did the rates of follow-up appointments: 58% and 80%, respectively. The results of the new research were reported at the Association for Research in Vision and Ophthalmology (ARVO) 2024 Annual Meeting.

The American Diabetes Association acknowledged in a 2017 position statement that retinal photography has the potential to bring screening into settings where optometrists or ophthalmologists are unavailable. This study shows the potential may not yet be realized. One key difference, study leader Madelyn Class, a medical student at Temple, said, was that the specialty clinics used a photographer trained in capturing ophthalmic images, while the primary care sites had medical assistants taking the photos.  “It seems user error played a role in the quality of photographs that were taken,” Class said. “Apparently, we will have to continuously monitor the performance of each photographer to ensure that quality photos are being taken.”

The findings may also point to the need for using different equipment for screening in primary care, Class added. “Robotic as opposed to manual cameras may help eliminate some of the user error that was experienced with primary care screenings,” she said.

Edited by Miriam Kaplan, PhD

Source: Richard Mark Kirkner, Medscape Medical News, May 22, 2024; see source article