EyeWorld is the official news magazine of the American Society of Cataract & Refractive Surgery.
Issue link: https://digital.eyeworld.org/i/906004
health industry. In the U.K., Google's DeepMind has been in collaboration with the National Health Service. Most recently it partnered with Moorfields Eye Hospital (London, U.K.) to see if its algorithms can pick up early signs of age-related macular degeneration (AMD) using 1 mil- lion anonymous optical coherence tomography (OCT) scans, according to a report by Business Insider. Tech companies aren't the only ones studying the use of machine learning systems in medicine. In January, a group of researchers at the Zhongshan Ophthalmic Centre, Sun Yat-sen University, Guangzhou published a study in Nature Biomedi- cal Engineering that demonstrated an AI system for identifying congenital cataracts. The machine's algorithm was able to diagnose the disease with more than 90% accuracy. 1 Dr. Abramoff's own research focuses on using AI for automated diagnostics to be used on the front lines of care, and he founded a com- pany (IDx LLC, Iowa City, Iowa), to commercialize this use of AI. "We build devices that analyze images from the retina for hemor- rhages and other abnormalities and decide that these patients are likely to have diabetic retinopathy from the images, without a human evalu- ating the images," he said. 2 Dr. Abramoff's artificial intelli- gence device is available for use in Europe and has completed clinical testing, and he is working with the FDA toward clearance in the U.S. "We hope that someday, just like in Europe, these systems will be in primary care and find those people who have diabetic retinopathy and are at risk for going blind," he said. Most of the work in ophthalmol- ogy has centered on computer-based image analysis, according to Dr. Chiang. "The rationale is it allows oph- thalmic diagnosis based on what is essentially visual pattern recogni- tion," he explained. Fit for ophthalmology Dr. Chiang's research in this area focuses on retinopathy of prematu- rity (ROP), one of the most common causes of childhood blindness in the world. "The problem is we as experts are often not consistent about diagnos- ing a baby who has 'plus disease' [the key marker of severe ROP]. Studies have shown that there will be people who look at the same retina and come up with a different diagnosis because the way that we diagnose is qualitative. Machine learning sys- tems can help make a better diagno- sis and a more consistent diagnosis," he said. Several groups have done work in ROP, trying to build systems that help experts make a better diagnosis, Dr. Chiang said. "We have developed computer systems to analyze images and make a diagnosis of plus disease and ROP, and we've shown that these comput- er systems perform comparably and often better than most experts for diagnosing this disease. That part is compelling. The computer systems are objective and reproducible." Much of the artificial intelligence work in medicine has been focused on diabetic retinopathy, Dr. Chiang said. In diabetic retinopathy, like several diseases in retina, there are very specific classifications for level of disease and how you treat it, which makes it more straightfor- ward to build computer systems, he explained. In a similar vein, Dr. Abramoff said, "I would argue that ophthal- mology is special because we have so many clinical studies especially about diabetic retinopathy that establish December 2017 • Ophthalmology Business 21 Detecting disease: Dr. Abramoff's artificial intelligence algorithm vs. a black box algorithm Source: Michael Abramoff, MD continued on page 22