Segmentation of white matter, gray matter, and CSF from MR brain images and extraction of vertebrae from MR spinal images

In this dissertation, we address two kinds of the biomedical images/volumes segmentation problems: (1) Segmentation of white matter (WM), gray matter (GM), and cerebral spinal fluid (CSF) from MR brain images/volumes; (2) Extraction of the vertebrae from MR spinal images. The former problem was notorious for the intensity inhomogeneity.; We propose a statistical decision model under a maximum a posterior probability (MAP) estimation and Markov random field (MRF) framework to segment MR brain images, where the spatial-varying Gaussian mixture (SVGM) is used to represent the intensity probability distribution of each of the three brain tissues (WM, GM and CSF), and MRF is used to estimate the prior probability. Three methods, the supervised method, the automatic (unsupervised) method, and the 3D method, are proposed to achieve the final segmentation.; The supervised method is a 2D method proposed to effectively learn the expert's segmentation and pursue the tissue labeling on 3D MR brain images with severe intensity non-uniformity. It consists of a parameter estimation process based on the modified kmean clustering (MKM) algorithm to estimate the parameters of SVGM from a radiologist-segmented image, and a classification algorithm based on iterated conditional modes (ICM) to perform the segmentation of the sequential brain images using the estimated parameters. Furthermore, this method can automatically update the parameters of SVGM from the most immediate practice. The excellent experimental results on 10 simulated and 20 in vivo MR brain volumes illustrate the efficiency of this method. Quantitatively comparing with other published methods, we show that SVGM-MRF is a potentially valuable method for the segmentation of high field MR brain images contaminated by severe intensity inhomogeneity.; The fully automatic or unsupervised 2D method based on the SVGM-MRF model is presented to accurately segment WM, GM, and CSF. The reference brain volume is obtained from a digital brain atlas instead of the radiologist segmentation. The parameters of SVGM are estimated from the reference images using either the expectation maximization (EM) algorithm or the MKM algorithm, and the final tissue labeling are obtained by using the ICM algorithm. The parameter estimation and the tissue labeling algorithms are iteratively operated until a stop criterion is satisfied. Furthermore we impose a criterion that minimizes the intensity means difference within the same segmentation tissue to improve the accuracy of the parameters of SVGM. (Abstract shortened by UMI.)...