Segmentation of brain tissues from neurological magnetic resonance (MR) images is a necessary and non-trivial procedure for volume measurement, three dimensional (3D) display, and feature analysis. These three techniques have clinical applications in diagnosis of disorders, such as Alzheimer's disease, epilepsy, atrophy, schizophrenia, and multiple sclerosis. An automatic procedure is clinically demanded for consistency of data analysis and reduction of processing time.
A procedure for automatic segmentation of brain tissues from multispectral MR images is developed. The automatic procedure consists of the following steps: (1) stripping away image pixels which represent skull and scalp; (2) correcting for image-intensity variation of a same tissue type across field-of-view (FOV) induced by radiofrequency (RF) inhomogeneity; (3) estimating model parameters of image samples for different tissue types (or training samples); and (4) segmenting the tissue types across the FOV. The procedure was tested by two sets of transaxial images acquired as relaxation time T1, T2, and proton density weighted by a 1.5 Tesla whole body MR scanner from patients. One set of images contains significant RF inhomogeneity and the other set has much less RF effect. The performance of the automatic procedure was validated quantitatively by comparing the segmented tissue regions with those labelled manually by an expert.
This talk will first introduce the basic concepts of MR image formation. The segmentation method will then be presented. Finally, the segmented results will be reported.