Every time I return from a scientific meeting, one or more of my patients is bound to ask me what the latest discoveries are in the field. One of the hottest topics at the recent 2018 Annual Meeting of the Association for Research in Vision and Ophthalmology Annual Meeting was artificial intelligence (AI) and its role in the automated detection and triage of patients with retinal disease. Our field readily lends itself to diagnostic approaches using imaging because retinal examinations are almost wholly based on visual inspection for ocular pathology. Machine learning and deep learning methods have developed early algorithms for automated evaluation of both diabetic retinopathy and age-related macular degeneration on fundus photographs.
Since the first large-scale validation of a deep learning algorithm for diabetic retinopathy was reported in JAMA in 2016, several groups have had their findings published in JAMA, JAMA Ophthalmology, and other journals, describing similarly successful efforts for automated detection of retinal diseases such as age-related macular degeneration.3 Major industry players, such as Google and IBM, have been quick to publicize their AI initiatives for the detection of diabetic eye disease. The application of AI to the evaluation and triage of retinal pathology hit a new milestone this April when the United States Food and Drug Administration granted approval for the first AI-based system for autonomous detection of diabetic retinopathy.
But what does the future hold for AI and the eye? Right now, most groups seem to be focused on identifying presence of disease or thresholds of referable retinal pathology. These efforts are already clinically applicable, particularly for large-scale screening programs to triage individuals who may be at risk for vision loss to more immediate access to specialized treatments and care. However, another promise of neural networks and their deployment in retinal disease ultimately may be the ability to identify novel features on retinal images that are able to predict outcomes and drive treatment decisions. Using deep learning methods, such factors already have been identified on retinal photographs that appear to predict age and sex with a remarkably high degree of precision.
A number of considerations will determine our future success in incorporating AI algorithms into clinical care. One of the current difficulties in developing these automated approaches is the sheer number of images needed to train and then validate these algorithms. In the midst of amassing enormously large image datasets that can number in the hundreds of thousands, it is critically important for AI groups to understand the importance of rigorous phenotyping of their images so that the algorithms developed are accurately trained to detect retinal disease. It would be a tragic waste of resources to train an AI system to recognize retinal disease based on sloppy or inaccurate image grading. As these methods become more sophisticated and we move into the arena of predicting outcomes after treatment or in the natural course of disease, it will be similarly important to use image sets that are associated with high-quality longitudinal outcome data. The most valuable data sets for these purposes are more likely to come from large-scale, prospective clinical studies, which mandate standardized methods for visual acuity and other outcomes assessments, rather than from standard care retinal clinics that vary more widely in their methods.
Until a couple of years ago, the intersection of AI and the detection and triage of retinal diseases seemed like something out of a science fiction story. But the future has arrived more quickly than many of us anticipated. In this new age, I don’t think that man will be supplanted by machines, but I do hold hope that our efforts in the restoration and preservation of vision worldwide will be made increasingly effective through AI technology.