Multimodal segmentation of brain MR images through hidden Markov random fields

Abstract

Segmentation of brain MR images, especially into three main tissue types: CSF, GM and WM, is an essential task in clinical applications as it aids surgical planning, computer-aided neurosurgery and diagnosis. However, every single MR image contains degenerative components such as noise and RF inhomogeneity which dramatically reduces the accuracy of the results of automatic post-processing techniques. A number of methods are proposed in the literature for tissue segmentation of brain MR images. Among these, Otsu thresholding, ML estimation and MRF model based methods are the ones that widely used. Moreover, 2D segmentation of True-T1 and True-T2 images almost completely removes the artifacts mentioned above hence results in the best results ever reported. However, the required scan time of the method and the expence of the process makes this method inapplicable to clinical applications. In this study, three di erent segmentation schemes for brain MR images, namely Otsu thresholding, ML classi cation and MRF model based segmentation, are analyzed taking the segmentation results of 2D segmented true parameter images and a novel multivariate MRF segmentation method using T1 and T2-weighted images is proposed. As a result, the performance of the segmentation methods when two dimensional data were used increased. Moreover, multivariate HMRF model-based segmentation method achieved the best results.|Keywords: Magnetic resonance imaging, Otsu thresholding, ML classification, MRF theory, multivariate segmentation.