Measurement of brain compartment volumes from MRI data using region growing and mixed volume methods

MRI data were collected from normal control subjects and from patients having AIDS or closed head injury (CHI). Both Tl-weighted and T2-weighted images were collected; in some cases transverse images were made, while in others coronal images were collected. The images were analyzed in three ways, each of which determined white matter (WM), gray matter(GM), and CSF volume, and the results were compared. Segmentation by interactive threshold selection or manual outlining is still the most commonly used method for compartment volume analysis. This method was applied to the control and AIDS brains, with repetition by two trained observers. These results were compared with those from a second method in which each voxel is viewed statistically as a mixture of whichever two compartments are closest at that anatomical location. Automatic thresholding, followed by regional skeletonization, is used to identify a small set of pixels in each slice that represent the central portion of each of the three regions of interest. All pixels in the slice are then subjected to an interpolation procedure in which their fractional composition is determined from these three intensity values. The intensities from the Tl-weighted images show contrast between GM, WM, and CSF, and these are used for the volume computation. Geometric information from the T2-weighted image is used for the location of the authentic compartment locations, as these images show strong contrast between the CSF and the skull. The third method studied is based on a three-dimensional region-growing algorithm which estimates each volume compartment by growing a volume from a seed point, limiting the growth based on spatial and feature criteria. The feature bounds are set restrictively so that GM, WM, and CSF regions are not contiguous, leaving a volume of mixed voxels between the regions. The distributions of intensities in each region are then used to interpolate the most likely composition of the unassigned voxels, so that volume mixing is assumed only in the spaces between the assigned regions. This method is quite robust, requiring little operator judgment. The volumes obtained by these methods are not substantially different, and the methods differ primarily with respect to interoperator variability and convenience of use. The third method also differs from the others in that it treats the set of slices as a single 3-dimensional data set, making better use of regional information. All methods reveal significant changes in brain compartment volume in cases of CNS pathology.