Purpose:
The built–in classifier (NFI) of the GDx VCC (Carl Zeiss Meditec, Inc., Dublin, CA) that classifies scans as healthy or glaucomatous on a scale of 1–100 is a linear support vector (SV) machine that depends on a large number (87) of features derived from the scans. A classifier based on only a few features may be more intuitive to the clinician, without adversely affecting the classification results. This study compares the cross–validated accuracy of a linear SV classifier based on many features with one based on only a few features for the classification of SLP images.
Methods:
71 healthy and 71 glaucomatous eyes were included. The optimal cut–off value and the corresponding accuracy of the built–in classifier were determined. A linear SV classifier based on all 87 features, thereby mimicking the built–in classifier, was trained and tested. A linear SV classifier based on a limited number of features (1–6) was trained and tested. All errors were estimated by cross–validation to avoid biased results.
Results:
The optimal cut–off value for the NFI was 42, resulting in an error of 9.2%. A linear SV classifier, trained on all 87 features of the available data, gave an error of 5.0% (SD 0.5%). Reducing the number of features to four hardly increased the error; the estimation of the error of this classifier was 5.9% (SD 2.5%). All results of the limited feature sets are tabulated below.
Conclusions:
Comparing the errors of the retrained SV classifier and the four–feature SV classifier shows no significant difference, although the error of the 87–feature classifier has a somewhat smaller variance. Adding more features to the four–feature classifier did not significantly reduce the estimated error of 5.9%.
Keywords: image processing • nerve fiber layer