Purchase this article with an account.
K. A. Townsend, G. Wollstein, D. Danks, K. Sung, H. Ishikawa, L. Kagemann, M. L. Gabriele, J. S. Schuman; Heidelberg Retina Tomography 3 Machine Learning Classifiers for Glaucoma Detection. Invest. Ophthalmol. Vis. Sci. 2007;48(13):3317.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Machine learning classifiers are computerized systems trained to detect the relationship between input parameters and a diagnosis. The purpose of this study was to assess the performance of classifiers trained on optic disc parameters obtained by Heidelberg Retina Tomography 3 (HRT 3; Heidelberg Engineering, Heidelberg, Germany) for discriminating between healthy and glaucomatous eyes. The performance of the classifiers was compared to currently available HRT 3 discrimination methods.
1 eye in 51 healthy subjects and 89 glaucoma subjects with good quality HRT scans was enrolled in the study. Classifiers were trained to discriminate between glaucomatous and healthy eyes using optic disc parameters derived from HRT 3 output. The classifiers were trained on 95 variables, with backward elimination in order to use the minimal number of parameters. Seven types of machine learning classifiers were trained on the parameters: Generalized Linear Model with Gaussian error, Generalized Linear Model with binomial error, Linear Discriminant Analysis, Support Vector Machine with linear kernel, Support Vector Machine with radial basis function (SVM-radial), Generalized Additive Model, and Recursive Partitioning and Regression Trees (RPART). The cross-validated area under the receiver operating characteristic curve (AUC) was calculated for all classifiers, individual parameters and the glaucoma probability score (GPS) as provided by HRT 3 software. AUCs of the machine classifiers were compared to the highest individual parameters and GPS scores using the DeLong method.
Visual field mean deviation was -0.64±;1.34dB for healthy subjects and -6.17±;6.92dB for glaucomatous subjects. The highest AUC for an individual parameter was 0.882 for the vertical cup/disc ratio and the global GPS AUC was 0.862. RPART with all parameters and with an optimized 10 parameter set provided significant improvement over global GPS (p=0.002, 0.006) and vertical cup/disk ratio (p=0.003, 0.009) with AUCs of 0.947 and 0.939 for each classifier respectively, followed by SVM-radial with an optimized 10 parameter set with significant improvement over vertical cup/disk ratio with an AUC of 0.916 (p=0.043).
Machine learning classifiers of HRT data may provide significant enhancement of the technology for detection of glaucomatous abnormality.
This PDF is available to Subscribers Only