Development of automatic techniques for segmentation of brain tissues from multispectral MR images

Automatic segmentation of brain tissues from multispectral magnetic resonance (MR) images (acquired as relaxation time T₁, T₂, and proton density P_D weighted by selecting appropriate settings of T_E and T_R, the spin-echo delay time and repetition time of a pulse sequence) requires automating the following steps: (1) compensation for image-intensity variation of a same tissue type induced by radiofrequency inhomogeneity across field-of-view; (2) stripping away image pixels which represent skull and scalp; and (3) estimation of model parameters (or training samples) such as image-intensity mean and variance for each tissue type and correlation coefficients for that tissue type among the three spatially-registered images (acquired in a very short time period). This proposed automatic approach first strips away the pixels of skull and scalp. The stripping is performed by detecting radially the pixels of inner-skull margin from the T₂ weighted transaxial image and applying the detected margin to remove the pixels of skull and scalp in the T₁ and P_D weighted transaxial images as well. The approach then determines the intensity-variation map within the margin from each stripped image and compensates for the variation by dividing that image by the determined map. Finally the approach estimates the model parameters by fitting the image data into a multivariate mixture. The fitting is performed by a maximum-likelihood estimator. The above three steps have been successfully implemented by a computer. The automatic approach was tested by a set of three MR images acquired by a 1.5 Tesla whole body scanner from a head. The removal of skull and scalp was very satisfactory. The compensation for the intensity variation improved significantly the estimation of model parameters