Contributions to the Automated Segmentation of Brain Tumors in Magnetic Resonance Images

This thesis studies the problem of the segmentation of magnetic resonance images (MRI) in patients with meningiomas and low grade gliomas. The studies are motivated by the potential of computer assisted neurosurgery to improve treatment outcome. To make such methods clinically practical, these techniques require the development of automated segmentation methods. First, the MR imaging characteristics of meningiomas and low grade gliomas are analyzed to assess the possibilities of automated segmentation. The analysis demonstrates that segmentation is not possible with a) statistical classification due to overlapping intensity distributions of tissue classes, or b) the spatial alignment (registration) of an anatomical normal brain atlas because such an atlas does not describe pathology. Subsequently, a segmentation framework that iterates statistical classification, local segmentation techniques and registration of a brain atlas is described and shown to allow complete segmentation of the skin surface, the brain, the ventricles and the tumor. A validation study with clinical MRI data demonstrates that the algorithm performs well despite the presence of overlapping intensity distributions of tissue classes. The reduction of operator time from 3 hours to 5-10 minutes makes it practical to consider the integration of computerized segmentation into clinical routine. The use of optical flow methods for the task of aligning an anatomical atlas to individual patient datasets is analyzed. An existing 2D optical flow approach that models discontinuous deformation fields is extended to 3D, and a novel N-channel probabilistic (NP) similarity measure is proposed that separates labels of different objects into different channels and incorporates classification probabilities as a confidence measure. A validation study with clinical and simulated MRI demonstrates that registration accuracy can be significantly improved for the alignment of multiple templates, and for the registration of templates to images with classification errors. The algorithms developed in this thesis are also used to project structural information (e.g. white matter tracts) from a normal brain atlas onto patients with brain tumors. The method provides anatomical information, which is not available from conventional MRI, for the planning of surgical approaches. This thesis contains three major contributions: a robust, reproducible and accurate method for the segmentation of brain tumors in MRI, a 3D adaptive regularization optical flow method and an N-channel probabilistic similarity measure for template matching.