Purchase this article with an account.
Xiaoqin Huang, Sidharth Mahotra, Tobias Elze, Mengyu Wang, Michael V. Boland, Louis Pasquale, Juleke Eugenie Anne Majoor, Hans Lemij, Kouros Nouri-Mahdavi, Chris A Johnson, Siamak Yousefi; Detection of Glaucoma Progression from Retinal Nerve Fiber Layer Thickness Measurements Using Machine Learning. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1005.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a machine learning model for detecting glaucoma progression from retinal nerve fiber layer (RNFL) thickness measurements acquired with optical coherence tomography (OCT).
We developed an unsupervised Gaussian mixture model with an expectation maximization (GEM) framework to identify clusters with similar RNFL thickness patterns in circular OCT scans (768 A-scans) from 691 eyes of 691 patients. We then identified the top prevalent eigenvectors (patterns) of the clusters representing eyes in the mild and advanced stages of glaucoma. We selected the slopes across these eigenvectors such that 95% of 2540 eyes in a separate longitudinal stable dataset (~9 follow-up visits) were not progressing. We used the selected slopes to test the model on a third longitudinal dataset with 254 eyes (~9 visits) and compared the detection rate of the proposed machine learning model with linear regression of RNFL summary parameters.
Machine learning discovered three clusters within 691 RNFL thickness measurements of 691 eyes (Fig. 1). A total of 12 patterns of RNFL loss (eigenvectors) were prevalent and accounted for about 77% of the total variability in the data (Fig. 2). The machine learning model detected glaucoma progression in 38.6% of the eyes tested, while average RNFL in global, superior, and inferior hemifields detected progression in 15.8%,13.0% and 15.0%, of the eyes, respectively.
The proposed machine learning construct identified a higher proportion of structural progression based on RNFL thickness measurements than linear regression of summary parameters. Additionally, the model identifies the major patterns of RNFL loss, which may aid clinicians in better monitoring of glaucoma subjects.
This is a 2021 ARVO Annual Meeting abstract.
Figure 1. Average RNFL thickness of three clusters that were identified using unsupervised Gaussian mixture model with expectation maximization (GEM) framework.
Figure 2. Patterns of RNFL loss identified using the machine learning approach.
This PDF is available to Subscribers Only