AUTOMATIC SEGMENTATION OF MEDICAL MR IMAGES

Abstract

Medical magnetic resonance (MR) images are rich in information content and are particularly useful for the analysis of soft tissues in the human body such as the brain tissues. The segmentation of the medical images is an indispensable step in medical image analysis and visualization. The segmentation can be performed automatically without user intervention, which has several benefits. So, the aim of this thesis is to develop effective and efficient methods for the automatic segmentation of the MRI images. In this thesis work, two effective methods are adopted and implemented for the automatic segmentation of the brain MR images into various tissue classes such as the white matter (WM), gray matter (GM), cerebro-spinal fluid (CSF), and the non-brain tissues. One method is based on the Fuzzy C-Means (FCM) algorithm; and the other one is based on the estimation of the maximum likelihood (ML) parameters of the Gaussian mixture model (GMM) using the expectation- maximization (EM) algorithm. The obtained results are analyzed and compared with the available reference "gold standards" or "ground truths". Also, two different ways of initialization of the class means for the above two methods are investigated. The genetic algorithm (GA) based initialization is found more reliable and efficient than the random way of initialization. A major problem for automated MR image segmentation is the corruption of the MR images with a smoothly varying intensity inhomogeneity or a bias field. For the bias field correction, the two step GMM-EM algorithm is extended to include a bias correction step. It improves the accuracy of segmentations effectively over the simple GMM-EM algorithm for the corrupted MR images.