Hi Jonathan,
Mixture modelling only assumes that there is a mixture distribution associated with the spatial maps. In general a Gaussian/Gamma MM (just like the Gaussian MM) is a universal approximator, i.e. it _can_ fit to any source density if the number of mixtures is sufficiently large. The Melodic MM is a Gaussian/Gamma MM with a _fixed_ number of mixtured (3) and a few assumptions that constrain the Gamma densities to only fit to the tails of the Gaussian distribution. This is more restrictive than a full flexible GMM but (i) makes this much more robust and (ii) reduces computational load very significantly (see our 2003 HBM poster) If signal is sparse and you use the appropriate normalisation steps then this model should still be fine, independent of the underlying data, i.e. it should work fine on e.g. structural data. Under these conditions it is only your (modality specific) SNR which determines how the 'non-background' tail voxels are relative to the bulk of the distribution - in your case you seem to have high SNR (relative to FMRI).
hth
Christian
On 29 Apr 2010, at 16:45, Jonathan O'Muircheartaigh wrote:
> Hi FSL experts,
>
> I've been using melodic to investigate a large sample of structural datasets, the results are as we would expect (and conform to an independent analysis) but I'm having trouble as to the post-stats mixture modelling part. The histograms outputted for each component are extremely sharp compared to what you would get in fMRI (see attached). Is mixture modelling appropriate in this instance, or is it robust to non-fMRI data?
>
> As always, apologies if I've missed the point.
>
> All the best,
> Jonathan
> <IC_24_MMfit.png>
|