Automated Patient-Specific Multi-tissue Segmentation of MR Images of the Head

The automated segmentation of magnetic resonance (MR) images of the human head is an active area of research in the field of neuroimaging. The resulting segmentation yields a patient-specific labeling of individual tissues and makes possible quantitative characterization of these tissues (e.g. in the study of Alzheimers disease and multiple sclerosis). The segmentation is also useful for assigning individual tissues conductivity or biomechanical properties for patient-specific electromagnetic and biomechanical simulations respectively. The former are of importance in applications such as EEG (electroencephalography) source localization in epilepsy patients and hyperthermia treatment planning for head and neck tumors. The latter are of interest in applications such as patient-specific motion correction and in surgical simulation. Automated and accurate segmentation of MR images is a challenging task in the field of neuroimaging because of noise, spatial intensity inhomogeneities, difficulty of MR intensity normalization and partial volume effects (a single voxel represents more than one tissue type). Consequently most of the techniques proposed to date require manual correction or intervention to achieve an accurate segmentation of the brain or whole-head. As a result they are time consuming, laborious and subjective. This thesis presents two automatic and unsupervised segmentation methods, for multi-tissue segmentation of the brain and whole-head respectively from multi-modal MR images, that are more accurate than the state-of-the-art algorithms. The brain segmentation method is based on the mean shift algorithm with a Bayesian-based adaptive bandwidth estimator. The method is called BAMS (Bayesian adaptive mean shift) and can be used to segment the brain into multiple tissue types; e.g. white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF). The accuracy of BAMS was evaluated relative to that of several competing methods using both synthetic and real MRI data. The results show that it is robust to both noise and spatial intensity inhomogeneities compared to competing methods. The whole-head segmentation method is based on a hierarchical segmentation approach (HSA) incorporating the BAMS method. The segmentation performance of HSA-BAMS was evaluated relative to a reference method BET-FAST (based on the BET and FAST tools in the well-known FMRIB Software Library) and three other instantiations of the HSA, using synthetic MRI data with varying noise levels, and real MRI data. The segmentation results show the efficacy and accuracy of proposed method and that it consistently outperforms the BETFAST reference method. HSA-BAMS was also evaluated indirectly in terms of its impact on the accuracy of EEG source localization using electromagnetic simulations based on a tissue conductivity labeling derived from the segmentation. The results demonstrate that HSA-BAMS outperforms the competing methods, and suggest that it has potential as a surrogate for manual segmentation for EEG source localization.