American Glaucoma Society Meeting

American Glaucoma Society Meeting

March 09, 2020
1 min read

Four AI models show high diagnostic accuracy for glaucoma

You've successfully added to your alerts. You will receive an email when new content is published.

Click Here to Manage Email Alerts

We were unable to process your request. Please try again later. If you continue to have this issue please contact

Grace Richter

WASHINGTON — Two conventional machine learning algorithms and two convolutional neural network models achieved high glaucoma diagnostic performances by reading full retinal nerve fiber layer thickness maps, according to a speaker here.

“All four machine learning models had [areas under the curve] of greater than 0.90, and they significantly outperformed mean circumpapillary RNFL thickness, which had an AUC of 0.76,” Grace Richter, MD, MPH, said at the American Glaucoma Society annual meeting.

Richter and colleagues assessed the diagnostic accuracy of multiple machine learning models using full RNFL thickness maps to detect glaucoma. The evaluation included 93 eyes of 69 subjects with glaucoma and 128 eyes of 128 healthy participants. The gold standard diagnosis of glaucoma was set by consensus from two fellowship-trained glaucoma specialists and a third specialist when disagreement arose. Each specialist evaluated clinical history, exam data, stereo disc photos and visual field results to determine the presence or absence of glaucoma, Richter said.

Researchers provided 6 mm by 6 mm, 200 pixel by 200 pixel RNFL thickness maps centered on the optic nerve head (Cirrus 4000, Zeiss) to four different machine learning algorithms. Two conventional machine learning models were used, including a k-nearest neighbor model, which compared test images with every training image to predict a label, and a support vector machine, which created a hyperplane in the feature space during training to perform classification, she said.

Two convolutional neural networks were also used, including GlaucomaNet, which was custom designed by the study authors, and ResNet-18, which was introduced in 2015 and is one of the most accurate algorithms to date, Richter said.

The k-nearest neighbor, support vector machine and ResNet-18 models all had diagnostic accuracy AUC of 0.91, while the GlaucomaNet model featured an AUC of 0.92.

“Machine learning of RNFL thickness maps could be applied to community glaucoma screenings in the future. Clinical should recognize the diagnostic value of assessing full RNFL thickness maps when evaluating for glaucoma,” Richter said. – by Robert Linnehan



Richter G. Machine learning models for diagnosing glaucoma from RNFL thickness maps. Presented at: American Glaucoma Society annual meeting; Feb. 27-March 1, 2020; Washington.

Disclosure: Richter reports she receives grant support from Zeiss.