AI useful in glaucoma screening
BOSTON – Artificial intelligence has enabled machines to accurately predict retinal nerve fiber layer thickness on OCT, Felipe A. Medeiros, MD, PhD, said here at the American Academy of Optometry annual meeting.
“Every day we see nerves that are suspicious, and we call them suspects,” he said. “It’s the majority of our patients – we can’t decide whether they have glaucoma or not.
“Why not have the deep learning model learn from the photos and predict the OCT result?” he continued. “It turns out the deep learning method does very well. There’s a very strong correlation between the predicted nerve fiber layer thickness from photos and the actual thickness you get on OCT.”
Medeiros said studies have shown the mean error to be about 7 microns, “which is quite small. It does a great job.”
He explained that the network accomplishes the prediction by looking at the area of the optic nerve and adjacent retinal nerve fiber layer.
“We’ve done this with minimum rim width, and, as you would expect, the results were quite similar,” Medeiros said. “Some have done it for predicting macular thickness measurements as well. We did this machine-to-machine model so we wouldn’t need the human labeling of photographs and have more specific cutoffs for screening. Also, we have recently shown that we can accurately predict change over time.”
Medeiros said that predicting progression was an unexpected finding.
Medeiros said he sees AI-based screening being applied to fundus photos and the technology being incorporated into pharmacies, similar to self-blood pressure testing. Potentially, the information could be added to the EHR, “and you could be flagged that you need to come in [for an exam].”
Glaucoma is difficult to diagnose on a single observation, he said, and such technology would enable patients to be monitored over time.
“The idea is to apply the photos more in the settings where you don’t have the opportunity to do OCT, but for those of us who have OCT, can AI help us in providing more information?” Medeiros asked.
“As you know, segmentation errors in OCT are really a problem when looking at scans for nerve fiber layer thickness, up to 40% sometimes,” he said. “They can go unnoticed unless you take the time to look at the B scan and see if the segmentation lines are correct.
“We developed deep learning that can help flag these segmentation errors,” Medeiros said.
The algorithm is trained on what to look for and can pick up errors that many would not notice, he said.
Collapse
Medeiros F. Artificial intelligence and glaucoma. Presented at: Academy 2021; Nov. 2-6, 2021; Boston.
Read more about