The purpose of the NMR segmentation tool is to evaluate methods for the segmention of tissue types in MR images and image sequences. Methods are provided for the analysis of single or multiple MR images. Density models of grey-level and image slope are constructed and used to compute the most likely voxel volume contents (Tina memo 2003-005).
The algorithm is a Bayes classifier which uses the parameters determined using an EM process (Tina memo 2001-005) (or from by the file provided) to compute the likely composition of each pixel (voxel). The method assumes that pure tissues have grey levels drawn from a Gaussian distribution and that partial volume effects are linear. These assumptions are reasonable for the majority of NMR scans in healthy subjects. The algorithm is stabilised (taking direct account of expected image noise) in the regions of pure tissue and guaranteed to give probabilities between 0 and 1, but may give systematic positional biases for some image types due to tissue boundary/noise ambiguity (Tina memo 2001-009).