Purchase this article with an account.
Michael Goldbaum, Siamak Yousefi, Akram Belghith, Linda Zangwill, Felipe Medeiros, Robert Weinreb, Daniel Meira-Freitas, Nima Hatami, Christopher Bowd; A Tree Classification Method for Identifying Normal Eyes, Non-Progressing Glaucoma Eyes, and Progressing Glaucoma Eyes from Spectral Domain OCT RNFL Thickness Measurements. Invest. Ophthalmol. Vis. Sci. 2013;54(15):4837.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To discriminate between normal eyes, non-progressing glaucoma eyes and progressing glaucoma eyes from longitudinal series of SD-OCT measurements (retinal nerve fiber layer [RNFL] thickness measured in 6 sectors and average RNFL thickness). We propose a binary classification to 1) discriminate normal from glaucoma eyes and 2) distinguish between non-progressing and progressing glaucoma eyes, using seven-dimensional RNFL measurements as input features. The final outcome assigned eyes to one of three classes: "Normal", "Non-progressing" or "Progressing".
A Bayesian classifier was trained on three separate Spectralis RNFL image series from normal eyes, non-progressing glaucoma eyes and progressing glaucoma eyes. The dataset included 73 images from 20 normal participants (mean follow-up 3 years, 3 tests), 331 images from 20 non-progressing glaucoma patients (imaged once a week for 5 consecutive weeks) and 81 images from 20 progressing glaucoma patients from the UCSD Diagnostic Innovations in Glaucoma Study (DIGS). Progression was defined by standardized assessment of stereophotographs and/or by designation as “likely progression” based on SAP Guided Progression Analysis (mean follow-up 4 years, 4 tests). 80% of each dataset was selected for classifier training and the remaining non-overlapping 20% was used for testing. First, images of normal eyes were separated from images of glaucoma eyes (non-progressing and progressing) then, non-progressors were separated from progressors. Classification was repeated 100 times and each time the samples were randomly permuted before dividing the dataset for training and testing. Confusion matrices representing specificity and sensitivity were computed.
Specificity for classifying all images from normal eyes was 80% and sensitivity for classifying all images from glaucoma eyes was 79%. Specificity for identifying images from non-progressing glaucoma eyes was 93% and sensitivity for identifying images from progressing glaucoma eyes was 80%.
Tree classification using a Bayesian strategy showed high specificity in normal eyes and high sensitivity in glaucoma eyes, with good accuracy for separating non-progressing from progressing glaucoma eyes.
This PDF is available to Subscribers Only