Diagonalized nearest neighbor pattern matching for brain tumor segmentation

Abstract

A new method is proposed for automatic recognition of brain tumors from MRI. The prevailing convention in the literature has been for humans to perform the recognition component of tumor segmentation, while computers automatically compute boundary delineation. This concept manifests as clinical tools where the user is required to select seed points or draw initial contours. The goal of this paper is to experiment with automating the recognition component of the image segmentation process. The main idea is to compute a map of the probability of pathology, and then segment this map instead of the original input intensity image. Alternatively, the map could be used as a feature channel in an existing tumor segmentation method. We compute our map by performing nearest neighbor pattern matching modified with our novel method of "diagonalization". Results are presented for a publicly available data set of brain tumors.