This article is more than 5 years old. Information may no longer be current.
Machine-to-machine learning may lead to improved grading of disc photos
Click Here to Manage Email Alerts
SAN FRANCISCO — A machine-to-machine method of deep learning may lead to a more accurate way to estimate the presence of glaucoma from photographs, according to a speaker.
“We ophthalmologists know that we perform very poorly in diagnosing glaucoma from photographs. It’s a challenging task unless you have very advanced disease. It’s a challenging task to look at a picture and determine whether the nerve has glaucoma or not,” Felipe A. Medeiros, MD, MPH, of Duke University, said at the American Glaucoma Society annual meeting.
Because ophthalmologists tend to underestimate or overestimate the amount of nerve damage seen in a photograph, a machine learning classifier or deep learning neuronetwork that is trained to replicate the gradings made by ophthalmologists would also replicate the errors made by ophthalmologists, Medeiros said.
Rather than training a neuronetwork to look at optic nerve photographs and predict the ophthalmologist’s subjective grading, Medeiros and colleagues undertook a study wherein the algorithm would predict the OCT’s objective findings of retinal nerve fiber layer thickness.
The study used 32,820 pairs of disc photographs of the optic nerve and OCT results from 2,312 eyes of 1,198 glaucoma suspects, participants with glaucoma or individuals with healthy eyes. The sample was divided 80% for training and 20% for testing.
The neuronetwork was trained to look at a photograph and predict the average nerve fiber layer thickness based on OCT input. In the test sample, the neuronetwork was asked to perform the prediction, and the value was compared with the actual OCT value to determine the accuracy of the program.
“Then we had the photos processed, and we trained this residual deep network, called ResNet34, to predict the global nerve fiber layer thickness. We also trained it to predict the OCT classification, like outside normal limits or borderline/normal,” Medeiros said.
Results of the test set included 6,292 pairs of photos and OCTs. Mean predicted nerve fiber layer thickness from the photographs was 83.3 µm, whereas mean observed nerve fiber layer thickness, which comes from the OCT of the same subject, was 82.5 µm.
“The machine-to-machine model does not require any human input for training the networks. It can provide a quantitative output that is not just a qualitative ‘yes or no glaucoma’ but a quantitative one,” Medeiros said.
The tool could potentially be used for longitudinal monitoring, although validation is still required, he said. – by Patricia Nale, ELS
Reference:
Medeiros FA. From machine to machine: An OCT trained deep learning algorithm for objective quantification of glaucomatous damage in optic disc photographs. Presented at: American Glaucoma Society annual meeting; March 14-17, 2019; San Francisco.
Disclosure: Medeiros reports he is a consultant for Allergan, Carl Zeiss Meditec, Novartis and Reichert, has received research support from Carl Zeiss Meditec, Heidelberg Engineering and Merck, and is cofounder of nGoggle Inc.