Kavya Kopparapu might be considered something of a “whiz kid.” After all, she had yet to enter her senior year of high school when she started Eyeagnosis, a smartphone app and 3D-printed lens that allows patients to be screened for diabetic retinopathy with a quick photo, avoiding the time and expense of a typical diagnostic procedure.
In June 2016, Kopparapu’s grandfather had recently been diagnosed with diabetic retinopathy, a complication of diabetes that damages retinal blood vessels and can eventually cause blindness. He caught the symptoms in time to receive treatment, but it was close. A little too close for Kopparapu’s comfort.
According to the IEEE Spectrum, Kopparapu, her 15-year-old brother Neeyanth and her classmate Justin Zhang trained an artificial intelligence system to scan photos of eyes and detect, and diagnose, signs of diabetic retinopathy. She unveiled the technology at the O’Reilly Artificial Intelligence conference in New York City in July.
After diving into internet-based research and emailing opthamologists, biochemists, epidemiologists, neuroscientists and the like, she and her team worked on the diagnostic AI using a machine-learning architecture called a convolutional neural network. CNNs, as they’re called, parse through vast data sets -- like photos -- to look for patterns of similarity, and to date have shown an aptitude for classifying images.
The network itself was the ResNet-50, developed by Microsoft. But to train it to make retinal diagnoses, Kopparapu had to feed it images from the National Institute of Health’s EyeGene database, which essentially taught the architecture how to spot signs of retinal degeneration.
One hospital has already tested the technology, fitting a 3D-printed lens onto a smartphone and training the phone’s flash to illuminate the retinas of five different patients. Tested against opthalmologists, the system went five for five on diagnoses.
Kopparapu’s invention still needs lots of tests and additional data to prove its efficacy before it sees widespread clinical adoption, but so far, it’s off to a pretty good start.
Eyeagnosis is operating in a space that's recently become interesting to some very large companies. Last fall, a team of Google researchers published a paper in the Journal of the American Medical Association showing that Google's deep learning algorithm, trained on a large data set of fundus images, can detect diabetic retinopathy with better than 90 percent accuracy.
That algorithm was then tested on 9,963 deidentified images retrospectively obtained from EyePACS in the United States, as well as three eye hospitals in India. A second, publicly available research data set of 1,748 was also used. The accuracy was determined by comparing its diagnoses to those done by a panel of at least seven U.S. board-certified ophthalmologists. The two data sets had 97.5 percent and 96.1 percent sensitivity, and 93.4 percent and 93.9 percent specificity respectively.
And Google isn’t the only player in that space. IBM has a technology utilizing a mix of deep learning, convolutional neural networks and visual analytics technology based on 35,000 images accessed via EyePACs; in research conducted earlier this year, the technology learned to identify lesions and other markers of damage to the retina’s blood vessels, collectively assessing the presence and severity of disease. In just 20 seconds, the method was successful in classifying diabetic retinopathy severity with 86 percent accuracy, suggesting doctors and clinicians could use the technology to have a better idea of how the disease progresses as well as identify effective treatment methods.
Lower-tech options are also taking a stab at improving access to screenings. Using a mix of in-office visits, telemedicine and web-based screening software, the Los Angeles Department of Health Services has been able to greatly expand the number of patients in its safety net hospital who got screenings and referrals. In an article published in the journal JAMA Internal Medicine, researchers describe how the two-year collaboration using Safety Net Connect’s eConsult platform resulted in more screenings, shorter wait times and fewer in-person specialty care visits. By deploying Safety Net Connect’s eConsult system to a group of 21,222 patients, the wait times for screens decreased by almost 90 percent, and overall screening rates for diabetic retinopathy increased 16 percent. The digital program also eliminated the need for 14,000 visits to specialty care professionals