Abstract We present an automatic segmentation method using the Maximum a posterior (MAP)-Markov random field (MRF) framework that possesses regional adaptive capability for the segmentation of MR brain images with the presence of noise and the intensity inhomogeneity. A spatial-varying Gaussian mixture (SVGM) is used to model the conditional probability distribution of each of the three brain tissues (WM, GM and CSF), and the MRF is used to represent prior probabilities. A three-component random vector consisting of spatial and intensity information is used in SVGM such that the model can represent more complex image characteristics, such as intensity inhomogeneity. The regional adaptive capability is achieved by imposing a similarity criterion that minimizes the intensity variation of the segmented tissues in local regions. Initial parameters of SVGM are estimated from a brain atlas using either the expectation maximization (EM) algorithm or modified k-mean (MKM) algorithm, and the iterated conditional modes (ICM) algorithm is used to perform the segmentation. The parameters estimation and the ICM algorithm are alternatively operated until a stopping criterion is reached. The implement of the method is validated by quantitatively comparing with other published classification methods on fifteen simulated and twenty in vivo brain volumes using the same evaluation criteria. Our algorithm demonstrates superior performance in all cases other than two, and is especially suitable for the volumes with higher noise and inhomogeneity. The complete experiments illustrate that the incorporations of the spatial information and the regional adaptive processing into MAP-MRF framework provide better solution to the segmentation of MR brain images. 1. Introduction The segmentation of the MR brain images into white matter (WM), gray matter (GM), and Cerebral Spinal Fluid (CSF) plays an exceptionally important role for numerous neuroanatomical analysis and applications [2][3]. While there are many research works done in the area, the segmentation of MR brain images is still a challenging unsolved problem due to the specific inherent difficulties of MR imaging [2], such as intensity inhomogeneity (non-uniformity or shading, more serious for higher fields [6][22]), partial volume effects [5], neuroanatomical structure variation, and random noise, and among which, the intensity inhomogeneity is the most serious one. This artifact is mainly derived from imperfect radio-frequency coils and the bulk magnetization susceptibility variability of tissues, and in some cases may produce more than 30% variations of image intensity for the same brain tissue [6], making the simple intensity-based classification methods fail to obtain good segmentation results. In the last decade, many approaches have been proposed to account for this artifact [1], [6], [8], [9], [11], [18], [20], [21], [22], [23] to improve segmentation results. First, the researchers’ efforts focused on how to correct the contaminated image before the segmentation, and then extract the tissue from corrected images assuming no inhomogeneity existence. The most intuitive ways are the use of a homomorphic filter and Phantom-based methods. Later, Likar [11] proposed a retrospective method based on the assumption that the entropy value of the inhomogeneity images would be larger than that of homogeneity images. Thus, they used a linear model to describe the inhomogeneity distribution, and the parameter values of the model were optimized using information minimization methods. Surface fitting methods are also widely utilized to overcome the intensity inhomogeneity. These methods try to compute a low order basis function (surface), such as a low-degree polynomial, to model the slowly varying bias field. Dawant et al. [8] proposed a surface fitting method to remove the inhomogeneity before image segmentation, in which a thin plate spline method was used to fit a surface to the intensity of the reference points selected either by users or by a classification algorithm. A quantitative comparison revealed that the results obtained from manually selected reference points were more reliable than those from computer selected points. Meyer et al. [6] employed a polynomial function to fit the bias field to regions that were segmented by LCJ (Liou–Chiu–Jain) segmentation algorithm [10]. The accuracy of the estimated intensity inhomogeneity depended on the precision of these prior Automatic Segmentation of MR Brain Images Using Spatial-Varying Gaussian Mixture and Markov Random Field Approach Zhigang Peng Department of Neurosurgery University of Illinois at Chicago Chicago, IL, 60612 William Wee Department of ECECS University of Cincinnati Cincinnati, OH, 45221-0030 Jing-Huei Lee Center for Imaging Research University of Cincinnati Cincinnati, OH, 45267-0586 Proceedings of the 2006 Conference on Computer Vision and Pattern Recognition Workshop (CVPRW’06) 0-7695-2646-2/06 $20.00 © 2006 IEEE