We describe our submission to the Brain Tumor Segmentation Challenge (BraTS) at MICCAI 2013. This segmentation approach is based on similarities between multi-channel patches. After patches are extracted from several MR channels for a test case, similar patches are found in training images for which label maps are known. These labels maps are then combined to result in a segmentation map for the test case. The labelling is performed, in a leave-one-out scheme, for each case of a publicly available training set, which consists of 30 real cases (20 high-grade gliomas, 10 low-grade gliomas) and 50 synthetic cases (25 high-grade gliomas, 25 low-grade gliomas). Promising results are shown on the training set, and we believe this algorithm would perform favourably well in comparison to the state of the art on a testing set.