Computational Techniques for Statistical Morphometric Analysis of 3-D MRI Data of Human Skull and Brain

Computational Techniques for Statistical Morphometric Analysis of 3-D MRI Data of Human Skull and Brain
Lin Shi

This thesis describes computational approaches to identifying statistical differentiations in anatomical structures in magnetic resonance images (MRI) between two groups of subjects, and to detecting the activation regions in functional MRI (fMRI). The efficacies of the proposed methodologies are demonstrated by detecting pathological differences in human skulls, brains, and vestibular systems for the etiology study of the adolescent idiopathic scoliosis (AIS).

A novel framework is proposed to detect morphological changes in the skull vault extracted from the T1-weighted MRI data. This framework includes 3-D skull vault segmentation, surface smoothing, distance-based feature extraction, and discriminative feature subset selection. The results of this study quantify the skull vault surface changes and point to the potential abnormality of bone growth regulation in AIS patients.

In particular, to automatically segment the skull vault with weak signal in T1-weighted MRI, we propose a model-based 3-D segmentation method. This method segments the skull vault using an active shape model, which is constructed based on the statistical information in skull vault surfaces segmented from CT data of a group of subjects. For the volumetric layers like the skull vault, the landmark correspondence is optimized by incorporating local radial thickness information.

Further analysis is performed toward detecting the statistical neuroanatomical differences. Volume-based morphometry methods have been applied to measure the statistical brain difference in T1-weighted MRI. To improve the accuracy of volume-based morphometry, we construct a brain template with 101 Hong Kong adolescent girls by iteratively using non-linear registration.

Besides, kernelized clustering methods, i.e., kernel K-means (KKM), spectral cluster analysis (SCA), and ellipsoidal support vector clustering (ESVC), are proposed to detect activation regions in functional MRI time series. KKM performs K-means clustering in the kernel space. SCA relies on spectral analysis of the similarity matrix to cluster fMRI time series. ESVC computes a compact hyper-ellipsoid enclosing the mapped features in the kernel space, and achieves a set of cluster boundaries in the original space.

Moreover, to test the hypothesis that unbalanced perception of the surroundings contributes to the spine deformity, we perform statistical surface-based morphometry on vestibular systems in AIS patients and matched normal controls. In order to obtain exact surface boundary of the vestibular system, we propose an automatic segmentation pipeline to delineate the vestibular system in T2-weigthed MRI images. Estimation of the overall labyrinth geometry is performed by measuring the relationships among the best-fit planes and the best-fit circles. Statistical shape differences are detected between AIS patients and normal controls.