Artificial intelligence has become a new player in health care and one that Neil Bressler, M.D., Wilmer’s James P. Gills Professor of Ophthalmology and chief of the Retina Division, sees as an ally in the battle to save the vision of people with age-related macular degeneration (AMD) and other retinal diseases. How? By diagnosing the cases early enough that intervention can succeed.
“There are eight million people who have an earlier, often asymptomatic stage of AMD that need careful monitoring, but we estimate only about four million of them know they have it,” says Bressler. He and his lab have teamed with the Johns Hopkins Applied Physics Lab to create Deep Convolutional Neural Networks (DCNNs), which are computer programs, to recognize the hallmark of AMD—drusen—on fundus images of the retina. Drusen are debris accumulating behind the retina, the light-sensitive tissue that lines the inside back wall of the eye.
DCNNs fall under the umbrella of machine learning, a field that seeks to teach computers to teach themselves new information using generic algorithms that can be applied to a wide range of data sets. What Bressler’s DCNN teaches itself is how to recognize drusen.
According to Bressler, the DCNN is told what the “ground truth is” by viewing images fed to it of fundus photographs of the retina—these images are referred to as the “training data set.” While viewing the training data set, the DCNN is told, ‘This fundus image has drusen, this one does not.’ After viewing thousands of sets of images, the program learns to sort the images into the two categories with increased accuracy.
Once the DCNN is trained, researchers feed it images without telling it, ‘This fundus image shows drusen, this one does not.’ Because the program has seen and sorted enough images in the training data set, it accurately recognizes images of drusen it has not seen before.
“What makes this artificial intelligence is that you’re not programming it and then it does its thing. You’re programming it to program or teach itself,” says Bressler.
Bressler’s lab recently published the results of its efforts, which found that the DCNN accurately identified AMD between 88.4 percent and 91.6 percent of the time, in JAMA (Journal of the American Medical Association) Ophthalmology. Such accuracy is comparable with human expert performance levels.
Once the DCNNs become as accurate as possible, then Bressler sees an opportunity for pieces of optical equipment—perhaps attachments to smartphones or kiosks in drug stores—to be created that can capture the fundus images in a cost-effective way.
“We don’t have the resources or the personnel to get to everybody and screen for asymptomatic, common retinal diseases that need monitoring or treatment,” says Bressler. “But if we can find very inexpensive ways to have the images analyzed to indicate, ‘This person has a high likelihood to need monitoring for AMD, but that person doesn’t,’ that has the potential to help an enormous number of people around the world.”