Training Computers to Combat Blindness
Web-connected cameras may help doctors detect a common eye disease.
Of all the complications of diabetes, few are as devastating as diabetic retinopathy, a progressive eye disease that causes blurred vision and in some patients, blindness. By the time most patients recognize something’s wrong, it’s often too late for them to be treated effectively. As a result, diabetes is the leading cause of vision loss among adults over 20. More than 12,000 new cases of blindness each year are caused by diabetic retinopathy, according to the National Institutes of Health.
An ophthalmologist and a scientist from the Department of Energy’s Oak Ridge National Laboratories in Tennessee believe they can help doctors detect diabetic retinopathy long before the disease wreaks havoc on their patients’ vision. Their startup company, Automated Medical Diagnostics (AMDx), has developed software that can detect the early signs of diabetic retinopathy by comparing digital photos of a patient’s retina to images that represent various stages of diabetic eye disease. AMDx’s founders believe their technology will enable all health workers–even those who are not trained in eye care–to take retinal scans of any patient, zap them over the Internet to AMDx’s servers, and get a diagnosis back before the patient leaves the office. “We’re trying to show we can be as accurate as a trained ophthalmologist,” says Ken Tobin, AMDx co-founder and division director of measurement science and systems engineering at Oak Ridge.
AMDx’s technology was inspired by a system that Oak Ridge scientists originally developed to help semiconductor manufacturers analyze defects in computer chips. Their software essentially teaches computers a technique called “content-based image retrieval.” The system can take a single image of a chip and then sort through giant databases of other images to find similar visual patterns–a process that some chipmakers now use to spot problems and improve manufacturing methods.
In 2005, Tobin met Edward Chaum, an ophthalmologist and professor at the University of Tennessee’s Hamilton Eye Institute in Memphis. “Less than half of diabetics are screened in any given year for retinopathy, despite the fact that they are told they need regular eye exams,” Chaum says. Many patients don’t have health insurance, he says, or they just don’t want the hassle of traveling to see yet another specialist. But Chaum and Tobin realized that if primary care doctors could do basic eye screenings on diabetic patients, they might catch many more cases of retinopathy than are being detected today.
The retina is particularly well suited to content-based image retrieval, Tobin says. Unlike other types of medical images, such as brain scans and mammograms–which are highly variable and require multiple images to create a three-dimensional effect–the retina is virtually two-dimensional and similar from one photo to another. That makes it easier for the technology to detect lesions, leaky blood vessels, swelling, and other abnormalities on the retina that can be early signs of disease. Chaum and Tobin spent five years developing algorithms that can extract information from retinal images and screen it against a database of more than 20,000 photos. AMDx doesn’t produce diagnoses but rather alerts doctors to patients who need to be referred to specialists for more in-depth testing, diagnoses and treatments.
AMDx is currently testing its system in a handful of clinics in Mississippi and Tennessee. Training doctors is easy, Chaum says, because the cameras have features like auto-focus, and they don’t require that the patients’ eyes be dilated. The photos are sent over the Internet to AMDx’s servers and automatically compared to images in its database. Chaum then checks each result manually–a process that takes about 90 seconds per case, he says.
AMDx’s goal is to ultimately turn over the entire job to its computers, but for now it must rely on Chaum’s review. That’s because some insurers–most notably Medicare and Medicaid–will only reimburse physicians for eye screenings after an ophthalmologist examines the results. Chaum and Tobin are collecting data, with the goal of proving to both regulators and insurers that their computers are as effective as Chaum at detecting disease. “The computers can handle thousands of reports a day. The bottleneck is me signing off on them,” Chaum admits.
Efforts to remotely diagnose eye diseases have been tried on a limited basis by the Veterans Administration hospitals and other institutions. Some ophthalmologists believe that if the idea catches on, they’ll be able to treat many more cases of diabetic retinopathy than they can today. “If you can teach a clinician to recognize changes in the eye, you can teach a computer to do it,” says Barrett Katz, an ophthalmologist and professor at Montefiore Medical Center and the Albert Einstein College of Medicine in the Bronx, NY. “It doesn’t take an ophthalmologist to gather these images–if anyone could do it, that would be a major step forward.”
Tobin and Chaum are on the hunt for venture capital to fund AMDx’s expansion. They hope the current focus on health reform will give them a boost, because the ongoing debate is drawing attention to the need for improving the efficiency of the health care system. “It’s not that we can’t treat diabetic retinopathy,” Chaum says. “It’s that we’re inefficient in how we screen for it.”
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today