Purchase this article with an account.
Mark Christopher, Akram Belghith, Christopher Bowd, Michael Henry Goldbaum, Luke J Saunders, Felipe Medeiros, Robert N Weinreb, Linda M Zangwill; Computational Features Derived from Retinal Nerve Fiber Layer (RNFL) Thickness Maps in Detecting and Monitoring Glaucoma. Invest. Ophthalmol. Vis. Sci. 2017;58(8):3999. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the use of computational features derived from RNFL maps in detecting primary open angle glaucoma (POAG) and identifying POAG progression.
Longitudinal data were collected from 179 eyes of 93 glaucomatous participants and 56 eyes of 28 heathy participants. These data included spectral-domain optical coherence tomography (SD-OCT), swept-source optical coherence tomography (SS-OCT), and standard automated perimetry (SAP) measurements collected regularly over a follow-up period averaging 1.9 years. SS-OCT wide-angle images were segmented using built-in software and manually reviewed for segmentation quality, resulting in a total of 1801 RNFL thickness maps. The thickness maps were used to identify a set of computational RNFL features using a principal component analysis (PCA)-based approach. The diagnostic accuracy of the top 5 features by explained variance was compared to the accuracy of average circumpapillary RNFL (cpRNFL) thickness and mean deviation (MD) of SAP 24-2 visual fields using area under receiver operating characteristic curve (AUC) at baseline. Progression was identified by applying mixed effect models to longitudinal data to separate age-related effects from glaucoma progression. A POAG eye was defined as progressing if the rate of change in structural or functional measurements was significantly different from zero (p<0.05) and faster than the 95th percentile of the healthy group.
The age-adjusted AUC of the computational RNFL thickness features was 0.94 for distinguishing POAG from healthy eyes at baseline, significantly greater (p<0.05) than both the 0.90 and 0.87 for cpRNFL thickness and MD, respectively. In detecting progression, computational features identified 9.5% of POAG eyes as progressing, while cpRNFL thickness identified 10.7% and MD identified 7.5%. The computational features detected 94% of progressing eyes defined using cpRNFL thickness and 93% of progressing eyes defined using MD.
The data-driven, PCA-based features derived from RNFL thickness maps offered improvements in POAG detection compared to cpRNFL and MD. These features also show promise in identifying POAG progression. The increasing availability of large OCT datasets allows data-driven techniques like these to discover novel imaging features that may offer advantages over many commonly used OCT-based measurements.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only