Feasibility of atlas-based segmentation of the brain in the presence of tumor by a weighted least-squares demons algorithm

Introduction MR-images of the brain can be segmented by registration with an atlas. A common way to do so is to first register the brain volume with an intensity atlas and consequently to propagate the labels of the atlas to the target brain volume according to the resulting deformation field. However, most intensity-based registration algorithms fail in the presence of pathologies, like brain tumors or MS lesions. The demons algorithm, as introduced by Thirion et al., starts from a constant brightness assumption, which means that intensities of corresponding voxels in the atlas and target image are assumed to be identical. In the presence of brain tumor, the intensities are altered and the constant brightness assumption is locally violated. As is pointed out by Li et al., this causes false deformations of the atlas and leads to gross segmentation errors in the vicinity of the tumor. We study the possibility to improve the robustness of the demons algorithm in the presence of pathology by minor, easy-to-implement, modifications of the demons force.