Nosarti, Chiara wrote:
> My understanding of segmentation in SPM2 is that it is based upon a
> three component normal mixture model with voxel intensities used as the
> feature variables.
Mostly right, but actually, from the paper
http://www.fil.ion.ucl.ac.uk/spm/doc/papers/john_multimodal.pdf
Generally, we use six or seven clusters: one each for
GM, WM, and CSF, two or three clusters to account for
scalp, eyes, etc., and a background cluster.
The code appears to have 8 Gaussians altogether (though I can't follow
it very well). SPM5, if you are interested, allows for multiple
Gaussians per tissue class (to model non-normality) and even lets the
user change how many are used.
> I would like to know whether the variances of the
> mixture components are allowed to vary.
I'm pretty sure they are, as in the usual GMM fitted by expectation
maximisation, with variable means, variances, and cluster parameters.
I think the subfunction update_cp_est in spm_segment performs a
standard EM update of (all) cluster parameter estimates. Though the
code isn't that well-commented -- sorry John, if you're reading ;-)
Best,
Ged.
|