Although nothing may ever totally replace the healing hands of a good physician, a new study reveals that artificial intelligence outperformed several ophthalmology experts in diagnosing a potentially blinding disease that affects premature infants.
Michael F Chiang, MD, a professor of Ophthalmology and Medical Informatics and Clinical Epidemiology at the Oregon Health and Science University (OHSU) School of Medicine, Portland, and a pediatric ophthalmologist at the Elks Children’s Eye Clinic, OHSU Casey Eye Institute, Portland, and co-author of the report, says the study shows that artificial intelligence may hold promise in aiding the diagnosis of retinopathy of prematurity (ROP) in babies.
“I hope pediatricians will take away that computer and information technologies are dramatically changing the practice of medicine, and that this requires that people from different clinical and scientific backgrounds work together,” Chiang says.
The study, published in JAMA Ophthalmology, evaluated 5511 retinal photographs using a new diagnostic algorithm with 5-fold cross-validation.1 The algorithm achieved 91% accuracy, outperforming 6 of 8 ROP experts.
According to the National Institutes of Health (NIH) National Eye Institute, ROP is a potentially blinding disease that primarily affects infants born weighing less than 2.75 pounds or before 31 weeks’ gestation. Other risk factors include anemia, blood transfusions, respiratory distress, and the overall health of the infant.
The condition develops when abnormal blood vessels grow and spread through the retina, and begin to leak and detach, scarring the retina and displacing it, according to the NIH. A common cause of vision loss in childhood, roughly half of the 28,000 children born each year in this population type are affected. Although some cases are mild and require no intervention, approximately 1500 of the infants born with ROP require medical treatment and 400 to 600 infants become legally blind as a result of the disease.
Retinopathy of prematurity is a leading cause of childhood blindness worldwide, according to the report, and is treated based on the presence of “plus disease,” a dilation and tortuosity of retinal vessels. Clinical diagnosis of this disease is subjective and variable, notes the report, and the goal of the study was to determine whether technology could aid in developing improved diagnostic tools.
Retinopathy of prematurity is traditionally diagnosed through an examination using dilation of the eyes and indirect ophthalmoscope in neonatal intensive care units (NICUs). “This is time intensive and logistically difficult for ophthalmologists, neonatologists, and NICU staff,” Chiang says. “Because of these challenges, telemedicine has become increasingly popular as an alternative method for ROP diagnosis in which retinal photographs are taken—often by NICU nurses—and transmitted securely to a remote ophthalmologist for diagnosis.”
The study was completed in collaboration with researchers at Massachusetts General Hospital (Boston), Northeastern University (Boston, Massachusetts), and the University of Illinois at Chicago, and demonstrates that computer-based diagnosis of “plus disease”—the key characteristic of severe treatment-requiring ROP—had comparable or better diagnostic accuracy compared with a group of 8 expert clinicians, Chiang says. Of the 5511 retinal photographs examined in the study, 82.3% were found to be normal, 14.6% were classified as pre-plus disease, and 3.1% were diagnosed as plus disease—results superior to those of human experts examining the patients.
“I hope computer-based tools like this will eventually improve the quality and delivery of ROP care,” Chiang says. “I believe this study is an example of how these tools can help provide diagnostic advice to ophthalmologists and neonatologists—and ultimately improve the quality of care for babies at risk for ROP.”
Standard treatments for retinopathy of prematurity include laser or pharmacologic therapy, but these interventions have adverse effects. More invasive surgical interventions may be required in more severe cases, such as those infants classified as having plus disease, wherein blood vessels become enlarged and twisted.
1. Brown JM, Campbell JP, Beers A, et al; Imaging and Informatics in Retinopathy of Prematurity (I-ROP) Research Consortium. Automated diagnosis of plus disease in retinopathy of prematurity using deep convolutional neural networks. JAMA Ophthalmol. May 2, 2018. Epub ahead of print.