Hi
On Wednesday, Oct 1, 2003, at 23:58 Europe/London, nima dehghani wrote:
> Hi
>
> i found that MELODIC, uses a guassian kernel for spatial smoothing!
>
> it seems to me that there are 2 problems with this:
>
> 1. ICA assumes that dara are NON-guassian....so, how we can use a
> gaussian kernel on a non-guassian data? is not it a fault?
>
Spatial smoothing is used to locally increase signal-to-noise. Any
pre-processing has an impact on the source distribution, but locally
convolving with a Gaussian kernel will *not* render the *entire*
spatial distribution being Gaussian - in fact, if 'signal' appears in
spatial clusters, spatial smoothing can indeed result in spatial
histograms with higher neg-entropy (i.e. in source distributions that
are 'easier' to extract due to increased non-Gaussianity) Smoothing too
much, however, is problematic (mainly for the mixture model inference
stage, though)
I generally advice to use very low amounts of smoothing (FWHM~ voxel
size). One of the nice advantages of the Mixture Model Inference stage
is, however, that (unlike GRF-Theory based inference) Gaussian
smoothing is no longer required. So if you're concerned about the
Gaussian smoothing, just switch it off or even use some more advanced
non-linear edge preserving technique like SUSAN.
> 2. when one tries to use ICA for multisubject analysis, spatial
> smoothing by a "unique kernel for all subjects" becomes a more serious
> problem!
>
Melodic does not yet include any multi-subject analysis options, I
suppose you're talking about concatenating data? What makes sense and
what doesn't will depend on what exactly you propose to do w.r.t.
multiple subjects. Per se, however, I can't see any serious problem
with a unique kernel - quite the opposite, in fact.
Maybe you would like to make your question more specific?
hope this makes sense
ta
christian
> do any of ICA experts havd an idea or a solution to this?
>
> am i right or not, with the 2 questions
>
> Regards
>
> Nima Dehghani MD.
>
> Schizophrenia, Cognition and Imaging Labaratoy; Department of
> Psychiatry; UBC; CA
>
> Email: [log in to unmask], [log in to unmask]
>
> --
>
> _______________________________________________
> Sign-up for your own personalized E-mail at Mail.com
>
> CareerBuilder.com has over 400,000 jobs. Be smarter about your job
> search
>
|