| This dissertation proposes novel automated segmentation methods for three dimensional (3D) Magnetic Resonance Microscopy (MRM) images of mouse brains. It consists of three major components as outlined below. First, a machine learning based segmentation using Support Vector Machine (SVM) is introduced to segment the 3D MRM images of five C57BL/6J mouse brains into 21 neuroanatomical structures. To apply SVM on segmenting mouse brain images, a multiclass imbalance problem needs to be handled. To address these issues, a mix-ratio sampling approach for SVM is proposed to determine various over-sampling ratios for different minority classes. Based on a comparison study, mix-ratio sampling method relieves the imbalance problem in multiclass more effectively and efficiently than the simple over-sampling method.;Secondly, a novel automated segmentation method, extended Markov Random Field (eMRF), is proposed for the mouse brain images. eMRF employs the posterior probability distribution obtained from SVM, which in general has a stronger discriminative power, to generate a classification based on the MR intensity information, uses a location prior for modeling location information and MRF for contextual information. To maximize the classification performance, eMRF uses the contribution weights optimally determined for each of the three kinds of information. Based on a comparison experiment, eMRF outperforms the three different existing methods: mix-ratio sampling SVM, the atlas based segmentation method and MRF.;Finally, another automated segmentation method, prior feature SVM-MRF (pSVMRF), is introduced to segment the mouse brain. The earlier work, extended MRF successfully showed that integration of SVM and MRF improves the segmentation performance compared with the existing methods. However, the computation of eMRF is very intensive, partly due to the computing demands of SVM training and testing. In this research, prior feature SVM (pSVM) is used to reduce the training and testing time of SVM and boost the classification performance. While the MR intensity information modeled by SVM and the location prior modeling the location information are linearly combined in eMRF, pSVM combines the two important pieces of information in a nonlinear fashion and, hence, enhances the discriminative ability of the algorithm. pSVMRF reduces the testing time substantially while improving the segmentation performance. |
