Segmenting brains

There are several reasons why a (reasonably good) segmentation can improve the analysis and visualization of your data. For instance, BrainVoyager QX and FreeSurfer support the matching of cortical brain surface representations (meshes) to an average group mesh, which means that the localization of functional activations (cluster peaks) can be usually performed with higher precision within the nomenclature of anatomical regions (not necessarily w.r.t. stereotactic coordinates of a template space). Another reason is that the visualization of results of a single subject on a surface (or group results on the surface representation of a representative subject) can improve the “readability” and comprehensiveness of those results.

While several segmentation algorithms exist, it is quite common that some amount of manual work has to be put into the segmentation to cope with the following problems (among others):

• local inhomogeneities and MRI noise can lead to a mis-labeling of voxels to either gray or white matter (or background)
• depending on the algorithm, holes in the segmentation (handles/bridges) occur
• spikes and other “unsmooth” areas remain (which lead to both reduced visualization quality but also certain problems when matching subjects)

The following steps are potentially beneficial for segmentation when working with anatomical datasets:

• inhomogeneity correction
• pre-labeling (for those voxels for which the label can be established with high to very high certainty)
• selective anatomical data smoothing (e.g. via sigma filtering or adaptive smoothing based on gradient information)
• high-resolution resampling (which allows a finer segmentation grid to be applied avoiding problems during mesh reconstruction)
• either by use of an algorithm (currently NOT implemented in NeuroElf) or manually segmenting the brain
• copying the pre-labeled voxels (to the high-resolution dataset)
• fine-tuning the segmentation