Influence of brain tissue segmentation on disease classifier accuracy
In this work we evaluate the impact of automated preprocessing of MR brain images on the prediction performance of a SVM AD classifier.Using a reference data set of 82 healthy controls and 82 individuals affected by AD we estimate the expected prediction accuracy and it’s confidence interval. The best model reaches an accuracy of 84.7± 2.2% when trained with 122 images. Best performance can be achieved when the input to the SVM is weighted by a custom computed tissue damage score map. We assess the usability of image characteristics metrics to detect severe classification or registration errors. No proposed metrics alone was a reliable indicator for general processing errors. Yet a sequence of thresholds applied to 3 different metrics discriminated regular segmentation and registration from non-regular processing with a sensitivity of 85% and a specificity of 97%, tested on a data set 481 images.