Magnetic Resonance Imaging (MRI) has become a popular method of high quality medical imaging. This is especially true in the brain where its non-intrusiveness is a definite advantage. With the increased usage, however, automatic segmentation of these non-trivial brain images has remained largely experimental. Our research has been geared towards solving this problem and earlier efforts [1,2,3] have shown that a combination of knowledge-based techniques and unsupervised fuzzy clustering could effectively detect slices with pathology and segment and label both a slice and partial volumes of a normal brain.
In this work, we describe the more difficult task of extracting tumor from slices found to have pathology by the systems in [1,2,3]. This is important because one of the uses of MRI data is tracking the size/shape of tumors as it responds (or doesn't) to treatment. Therefore, an automatic and successful method for segmenting tumor from the rest of pathology would be a useful tool. Of the many tumor types that are found in the brain, this work focuses glioblastoma multiformes. This tumor type was addressed first because of its relatively compact and well defined nature.