Purchase this article with an account.
M. A. Mayer, J. Hornegger, C. Y. Mardin, F. E. Kruse, R. P. Tornow; Automated Glaucoma Classification Using Nerve Fiber Layer Segmentations on Circular Spectral Domain OCT B-Scans. Invest. Ophthalmol. Vis. Sci. 2009;50(13):1101.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To inspect the possibility for a glaucoma classification using an automated nerve fiber layer segmentation on circular OCT scans.
Circular B-scans (diameter 3.4mm, 512 or 768 A-scans) around the optic disk were acquired from 204 subjects using a spectral domain OCT system (Spectralis HRA+OCT, Heidelberg Engineering). The patients were diagnosed by experts and separated into a Normal (N, 132 subjects) and Glaucoma (G, 72 subjects) group. This leads to a two-class classification problem. The method for an automated glaucoma classification was as follows: An automated nerve fiber layer (NFL) segmentation algorithm developed at our department was used to obtain NFL thickness profiles (example see figure). Out of these profiles the following two feature types were generated resulting in 14 features:- The minimum, maximum and mean was calculated for: All profile values, the one-third biggest and the one-third smalles ones.- The thickness profiles (768 and 512 A-Scans) were reduced to 128 values by averaging neighbours. This vector was further compressed to five values using principal component analysis.To eliminate the possibility that the age related degradation of the NFL affects the results, for the training and testing of the classifier subjects were excluded randomly in such a way that the age-distribution was equal in all age decades. This reduction left 61 N and 61 G patients. A Support Vector Machine classifier was used for classification. A 10 fold crossvalidation determined the result. The experiments were carried out 10 times to capture the variation of the different random exclusions.
The classification accuracy was on average 88,5% and the ROC area 88,5%, too. For glaucoma detection a sensitivity of 91,1% and a specificity of 87,0% was achieved.
The results show that the NFL thickness from a single circular B-Scan is a useful glaucoma indicator that allows automated classification with a high accuracy. A possible application of such a classification is to use it in cheap OCT devices for a mass glaucoma screening.
This PDF is available to Subscribers Only