Multimodal segmentation of brain MR images through hidden Markov random fields

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 nuerosurgery 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 most successful outcomes ever reported. However, the required scan time of the method and the expence of the process makes it inapplicable to clinical practices. In this study, three different segmentation schemes for brain MR images, namely Otsu thresholding, ML classification and MRF model based segmentation are analyzed taking the segmentation results of 2D segmented true parameter images as golden standards and a novel multivariate HMRF segmentation method using T₁ and T₂-weighted images is proposed.