Gaussian Neighborhood Descriptors for Brain Segmentation
Henrik Skibbe, Marco Reisert, Hans Burkhardt · 2011
In this paper we introduce a novel way for describing and classifying high angular resolution diffusionweighted magnetic-resonance images (HARDI) of the human brain. Our approach is capable to segment the brain images into gray matter (GM) and white matter (WM) tissue. For the segmentation a two step approach is suggested: The appearance of a training image is described locally at each voxel position in a rotation invariant manner. Then a classifier is trained and used for distinguishing between background (BG), GM and WM in unclassified images. In contrast to existing model-free methods we are not only using the raw measurements at each position, we also comprise neighboring measurements in a rotation invariant way. Experiments show that our method outperforms existing methods significantly. Furthermore, we show that our method gives also reasonable results for brains with pathologies like tumors. 1