Software - Image Segmentation
MRI tissues can be classified either fully automatically or semi-automatically. In the former case, no user interaction is needed or required. In the semi-automatic mode, the user has to input estimates of the mean intensity values for each tissue type via the gui interface.
The algorithm models the tissue distributions as Gaussians and uses a maximal-likelihood process to estimate the parameters (mean and standard deviation of each distribution) as well as the most likely tissue class to which each voxel should belong. It treats the array of voxels as a Markov random field, meaning here that the probabilities of a voxel belonging to a given class are influenced only by the classifications of neighboring voxels. Thus it is an EM-style algorithm in which each iteration first estimates the most likely Gaussian parameters based on the current voxel tissue classifications, and then the voxel classifications are adjusted to maximize their likelihood under the updated parameter estimates.
The algorithm iterates until the number of tissue reclassifications drops below a low threshold. This is usually no more than 12 iterations and takes a few minutes at most. The results are fairly good, though MRI field inhomogeneity can skew the segmentation, as can poor initial estimates of the tissue means.
Figure Segmenter images. Left window shows image histogram with tissue peaks selected as input to the algorithm. Right: image after segmentation.