Parameter Estimation and NMR Image Segmentation

A statistical method is proposed to classify tissue types and to segment the corresponding tissue regions from relaxation time T1,T2,and proton density NH weighted nuclear magnetic resonance images. The method assumes that the distribution of image intensities associated with each tissue type canbe expressed asa multivariate likelihood function of threevariables (T1,T2,NH)at each location witbin that tissue region. The method further assumes that the underlying piecewise contiguous tissueregions can be characteked by a Markov random field prior. In classifying the tissue types, themethod uses a finite multivariate mixture as the likelihood of realizing the images given the underlying finite number of tissue regions. The parameters associated with each tissue type (i.e., the mean, variance and correlation coeflicient of the multivariate function, and the number of voxels within each tissue region) are estimated by maximum likelihood fitting the image data via the expectation-maximization algorithm. The number of tissue types is determined by the information criterion of minimum description length. The method segments the tissue regions given the estimated tissue parameters by maximum a posteriori probability. The prior is constructed by the region membership of the first- and second-order neighborhood. The method is tested by a set of TI,T2, and NH weighted images of the brain acquired with a 1.5 Tesla whole body scanner. The number of tissue types and the parameters of these tissue types are satisfactorily estimated. The regions of dif€erent tissue types are successfully segmented.