Hi Jonathan,
I am not an FSL expert, so I should probably not reply to your question.
I apologise to everyone if I am being rude by pushing my personal work
on the mailing list of a project that I don't contribute to.
On Thu, Apr 29, 2010 at 04:45:50PM +0100, Jonathan O'Muircheartaigh
wrote:
> 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?
From looking at the histograms output when running Melodic, I would agree
with you that on the artifactual components, mixture modeling is
dangerous. My interpretation is that the model will fit a null
distribution to the data independent of what this data is, and as a
consequence has a tendency to overfit the null on artifact components. I
believe that the mixture model in Melodic has informative priors to
alievate this problem, but it seems to me that there is still some
overfitting going on.
To address this issue, we have developed a multivariate probabilistic
model consistent with ICA and based on the assumption of sparsity. This
model introduces naturally a thresholding criteria after the fit. It has
been presented at ISBI last month, but the proceedings have not yet been
published. I have uploaded a preprint to
http://gael-varoquauxinfo/varoquaux_isbi_2010.pdf.
Best,
Gaël
--
Gael Varoquaux
Research Fellow, INRIA
Laboratoire de Neuro-Imagerie Assistee par Ordinateur
NeuroSpin/CEA Saclay , Bat 145, 91191 Gif-sur-Yvette France
++ 33-1-69-08-78-35
++ 33-6-28-25-64-62
http://gael-varoquaux.info
|