Unsupervised Segmentation of Head Tissues from Multi-Modal Magnetic Resonance Images: With Application to EEG Source Localization and Stroke Detection
The automated segmentation or labeling of individual tissues in magnetic resonance (MR) images of the human head is an essential first step in several biomedical applications. The resulting segmentation yields a patient-specific labeling of individual tissues that can be used to quantitatively characterize these tissues (e.g. in the study of Alzheimers disease and multiple sclerosis) or to assign individual dielectric properties for patient-specific electromagnetic simulations (e.g. in applications such as electroencephalography source localization in epilepsy patients and microwave imaging for stroke detection). Automated and accurate segmentation of MR images is a challenging task because of the complexity and variability of the underlying anatomy and the noise and the bias field (spatial intensity inhomogeneities). Consequently, manual segmentation, including both interactive segmentation and manual correction, is largely used in clinical research. However, it is time consuming, subjective, tedious, and labor-intensive. This thesis presents new segmentation methods for both the brain and whole-head that are both automatic and accurate. It also presents empirical evaluations of these methods both directly in terms of segmentation accuracy and indirectly in terms of efficacy in electroencephalography (EEG) source localization and stroke detection. The evaluations were performed using both synthetic and real MRI data. This thesis makes four distinct contributions. The first is a novel unsupervised segmentation frame- work for segmenting MR images of the brain into three tissue types: white matter, gray matter and cerebrospinal fluid. It is a combination of Bayesian-based adaptive mean shift, incorporating an a priori tissue label probability maps, and the fuzzy c-means algorithm. The experimental results —based on both synthetic T1-weighted MR images for different noise levels and spatial intensity inhomogeneity levels, and real T1-weighted MR images —demonstrate its robustness and that it has a higher degree of segmentation accuracy than existing methods. The second is a novel automated unsupervised whole-head segmentation method for the purpose of constructing a patient-specific dielectric or biomechanical head model. The method is based on a hierarchical segmentation approach incorporating Bayesian-based adaptive mean shift. The experimental results demonstrate the efficacy of the proposed method, its robustness to noise and the bias field, and that it has a higher degree of segmentation accuracy than existing methods. The third is an evaluation of the proposed whole-head segmentation method in the context of EEG source localization. The experimental results show that the proposed method yields improved localization accuracy over the commonly used method for constructing a realistic head conductivity model for EEG source localization. The fourth is an evaluation of several existing unsupervised segmentation methods including the proposed whole-head segmentation method in the context of stroke detection using a microwave imaging system. The experimental results show that the proposed method has higher image reconstruction accuracy for intracerebral hemorrhage compared to the existing methods. The results also suggest that accurate automated segmentation can be used as a surrogate for manual segmentation to obtain accurate image reconstruction of an intracerebral hemorrhage and can assist in real time stroke detection.
EEG source localization
Room EA, floor 4, Hörsalsv ̈agen 11, Chalmers University of Technology
Opponent: Dr. Ali R. Khan, Robarts Research Institute, Western University, London, Ontario, Canada