Simulation based analysis of automated, classification of medical images.

OBJECTIVES The ability of various classifiers to discriminate between normal and glaucomatous eyes based on features derived from automated analysis of laser scanning images of the eye background is investigated. METHODS To compare the classifiers without over-optimization for a given dataset, we use a simulation model to create topography images. We designed three different simulation setups as model of extreme situations and medical subgroups. RESULTS Neither linear nor tree-based classifiers are ideal for all setups. The most robust performance is obtained by a combination of both, so-called Double-Bagging. Classification of real data from a case-control study shows best results with Double-Bagging. All results obtained with the analysis method extracting features automatically are worse than those obtained by the same classifiers but with features derived from an analysis method that requires intervention of a physician. CONCLUSIONS Robust classification results for classification of laser scanning images obtained with the Heidelberg Retina Tomograph are achieved by combined classifiers. The examined automated procedure causes an increased misclassification error compared to the established clinical routine requiring an expert physician's intervention.