Label propagation with robust initialization for brain tumor segmentation
Hongming Li, Yong Fan · 2012
A fully automatic segmentation algorithm based on local and global consistency with robust label initialization is proposed for brain tumor segmentation in multi-parametric MR images. A tumor probability map is first computed using support vector machine (SVM) classification of voxel-wise features, and then aggregated with respect to the image intrinsic structures in a multi-scale manner. Salient candidate tumor regions are extracted from the multi-scale hierarchy by a novel strategy based on perceptually important points. Robust label initialization is finally generated taking into account the information from both SVM classification and salient candidate regions. Validation experiment results on multi-parametric MR images have demonstrated that improved tumor segmentation accuracy can be achieved compared with state-to-the-art methods.