Purchase this article with an account.
D. Bizios, B. Bengtsson, J. L. Hougaard, A. Heijl; Processing of Optical Coherence Tomography (OCT) Data for Glaucoma Detection With Machine Learning Classifiers. Invest. Ophthalmol. Vis. Sci. 2007;48(13):525.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To test whether data reduction and feature extraction techniques can provide input to machine classifiers that is useful for discrimination between healthy individuals and glaucoma patients. Less complex data models with fewer, more robust parameters would facilitate data representation and analysis, and improve classification on limited sample sizes.
Two algorithms were used to process the output of 152 OCT tests, obtained with the Fast Retinal Nerve Fiber Layer (RNFL) protocol, from 90 healthy individuals and 62 glaucoma patients. These were the ISOMAP, and the algorithm based on Locally Linear Embedding (LLE). Both techniques are able to reduce the 256 A-scans of each OCT test into a smaller number of parameters, while preserving relevant information existing in the A-scan measurements. The A-scans from each OCT test were reduced to 5 parameters, which were then used as input to an Artificial Neural Network (ANN) and a Support Vector Machine (SVM). Classifier performance on the reduced data as well as on the 256 A-scans was measured by the area under the receiver operating characteristic (ROC) curve derived from 10-fold cross-validation.
For the ANN trained only on the 5 parameters, the area under the curve was 0.957 with the ISOMAP and 0.974 with the LLE algorithm. The SVM trained on the reduced data, achieved an area under the curve of 0.975 with the ISOMAP and 0.976 with the LLE algorithm. When using the 256 A-scans as input, the area under the curve was 0.974 for the ANN and 0.972 for the SVM. No statistically significant differences were found between the ROC curves.
Both algorithms were able to preserve relevant information while reducing the 256 A-scans of each OCT test into a few parameters. The effective representation of the A-scans by only 5 parameters provided a classification performance comparable to that of the full dataset. Performance advantages with these techniques should become more prominent with increasing number of test parameters in future generations of OCT instruments.
This PDF is available to Subscribers Only