Context-sensitive Classification Forests for Segmentation of Brain Tumor Tissues

We describe our submission to the Brain Tumor Segmentation Challenge (BraTS) at MICCAI 2012, which is based on our method for tissue-specific segmentation of high-grade brain tumors [3]. The main idea is to cast the segmentation as a classification task, and use the discriminative power of context information. We realize this idea by equipping a classification forest (CF) with spatially non-local features to represent the data, and by providing the CF with initial probability estimates for the single tissue classes as additional input (along-side the MRI channels). The initial probabilities are patient-specific, and computed at test time based on a learned model of intensity. Through the combination of the initial probabilities and the non-local features, our approach is able to capture the context information for each data point. Our method is fully automatic, with segmentation run times in the range of 1-2 minutes per patient. We evaluate the submission by cross-validation on the real and synthetic, high- and low-grade tumor BraTS data sets.