> It does not use a T1 or T2 image anymore. The priors that are used are
> determined by opts_seg.tpm and they default to gray/white/csf.nii in the
> tpm directory in spm (type 'help spm_preproc'). If the question was how
> spm determines the modality of the input image, I never checked that.
The initial registration is based on mutual information between the tissue
probability maps and the individual's image, and works without knowing the
type of sequence that was used. The segmentatation toolbox within SPM8 does
this step slightly differently (and is more robust), by effectively
incorporating a crude segmentation within the registration (as well as using
skull and scalp info).
After the initial affine registration, it is possible to roughly label which
voxels are GM, WM etc. With this approximation, the intensity distribution
of the different tissue types can be re-estimated. With more accurate
estimates of intensity distributions of the tissue types, the GM, WM etc can
be labelled again... ...and so on.
Improved labelling of GM, WM etc allows the tissue probability maps to be more
acurately overlayed (by nonlinear registration), which in turn allows more
accurate labelling.
It's all a bit circular, so involves lots of alternating among the different
steps - along with a bit of intensity nonuniformity correction for good
measure.
Best regards,
-John
|