In this paper, we consider the problem of fully automatic brain tumor segmentation in multimodal magnetic resonance images. In contrast to applying classification on entire volume data, which requires heavy load of both computation and memory, we propose a two-stage approach. We first normalize image intensity and segment the whole tumor by utilizing the anatomy structure information. By dilating the initial segmented tumor as the region of interest (ROI), we then employ the random forest classifier on the voxels, which lie in the ROI, for multi-class tumor segmentation. Followed by a novel pathology-guided refinement, some mislabels of random forest can be corrected. We report promising results obtained using BraTS 2015 training dataset.
展开▼
