Fully automatic skull stripping of routine clinical neurological NMR data

Image analysis of neurological NMR data is often an easier undertaking when non-cerebral tissue compartment voxels are removed from the NMR image dataset. This preprocessing step is often called ´skull stripping´. The simple but robust technique formulated and presented in this paper utilizes a combination of mathematical morphology and statistical segmentation techniques. Non-tissue background voxels are deemed to possess a Rayleigh distribution and consequently removed using an adaptive region dividing technique. Further processing automatically identifies a set of voxels that act as a test slice to determine whether the cerebral tissue compartment voxels have been fully separated during subsequent morphological processing. This set is used as a test to terminate an iterative morphological processing scheme to disconnect cerebral from non-cerebral voxels. The method has been successfully applied to 9 NMR datasets of varying quality with low inter-slice resolution. It therefore appears that this approach should be sufficiently robust to be useful for the statistical analysis of routine clinical NMR data.