Investigators used 500K images to hone the game-changing technology.
Image has been resized. Courtesy: Tmhlee, Wikimedia Commons.
When a team of international researchers began testing a new deep learning system (DLS), they hoped to gauge how well it could identify diabetic retinopathy and similar diseases. The technology, it turned out, had a keen eye.
In the primary data set, which took into account 71,896 images from 14,880 patients, the machine learning technology boasted a sensitivity of 90.50% and a specificity of 91.60%, according to the findings, published this week in JAMA. “The DLS has high sensitivity and specificity for identifying diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes,” the researchers wrote.
The takeaway? The artificial intelligence engine’s ability to spot diabetic retinopathy and other diseases—including vision-threatening diabetic retinopathy, possible glaucoma, and age-related macular degeneration—impressed investigators. They called for additional research to examine how healthcare providers and hospitals can leverage the technology to “improve vision outcomes,” according to the study.
In total, researchers fed roughly 495,000 retinal images from patients of diverse backgrounds to the DLS, training it to detect diabetic retinopathy and the other diseases, a process that concluded in May 2016. The photographs came from tens of thousands of patients, some of whom had 1 or more of the disease, and others who did not, according to the study.
Later, in May 2017, hundreds of thousands of additional photographs then tested the DLS’s capabilities. Researchers pitted the technology against graders, including retinal specialists, general ophthalmologists, and optometrists, hypothesizing that the computer would perform at least as well as the people. Ultimately, according to the findings, machine learning surpassed the graders in some areas and fell short in others.
“The performance of the DLS was comparable and clinically acceptable to the current model based on assessment of retinal images by trained professional graders and showed consistency in 10 external validation data sets,” the researchers wrote. What’s more, the DLS “showed consistent diagnostic performance across images of varying quality and different camera types, and across patients with varying systemic glycemic control level.”
The team also examined how the technology could be used in 2 “common” diabetic retinopathy screening models. One was fully automated and met performance standards for all 3 diseases, while a separate semi-automated model showed potential for DLS incorporation down the road, according to the study.
Despite the promising results, researchers noted that their work has limitations. The training set of images, for instance, was not fully crafted based on specialist grading. There also exists “black-box” issues that could affect clinical adoption.
This isn’t the first time that the research community has pointed artificial intelligence at diabetic retinopathy, recalling an earlier successful effort by Google. Similar initiatives have focused on other parts of the body. One such investigation this fall examined chest X-ray images, strengthening the technology and, researchers hope, diagnoses.
Investigators from the Singapore Eye Research Institute, the Duke-NUS Medical School, the Johns Hopkins Wilmer Eye Institute, and other academic centers across the globe contributed to the study. “Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes” was published yesterday in JAMA.