Hi Christina,
melodic.log is ceated by the melodic binary, so all the pre-
processing which is initiated by the melodic GUI is not recorded in
there but in report.log. There should be a command line call to $
{FSLDIR}/bin/ip - the first number after the -t option relates to the
highpass filter sigma (in volumes), i.e. if TR is 3 sec and you're
using a 100s hp filter then this number is 100/2*TR = 16.667.
Denoising is straight forward. First, run melodic on the original
data (filterd_func_data.nii.gz, say). This gives an .ica directory
(blah.ica, say). Go throught the components in there and note down
the numbers of the components which you are certain (!) are
structured noise. Then simply run
melodic -i filtered_func_data --ICs=blah.ica/melodic_IC --
mix=blah.ica/melodic_mix.txt -o blah2.ica
In blah2.ica you should then find a file melodic_ICAfiltered - which
is the original data (filtered_func_data.nii.gz) where the structured
noise effects have been removed using simple linear regression. Note
that this results in a small loss in degrees of freedom which - if
the original time series is reasonably long and if you have not
removed too many components - is pretty irrelevant for the final GLM
stats. We're working on integrating melodic denoising into feat,
which will then take care of these things automagically, watch this
space ;)
Wrt your final question: task related signal changes are treated like
any other form of signal change in the data. If these changes are
structured variations then ICA will try to represent these components
as components. If a task-paradigm is used then any induced variation
has no more (but also no less) impact on the remaining components
than any other component being estimated.
In theory, the existence of an additional source will simply increase
the estimate for the number of components so that the additional
signal will be represented as an additional source estimate.
hope this helps
christian
On 25 May 2007, at 12:23, Christina Hugenschmidt wrote:
> Hi!
>
> We are enjoying exploring ICA with melodic, and as you all, especially
> Christian, have been most helpful in answering our questions, I was
> hoping
> you might be kind enough to answer a few more.
>
> First, where can you find information about what high pass filter
> was used
> in an analysis? I (shamefully) did not write down the high pass
> filter used
> in a couple of analyses, and did not see it in the melodic log.
>
> Also, we have a subject whose data did not converge. In reading
> posts about
> this, I saw the suggestion that the data be denoised using melodic.
> The
> post mentioned the following link
> (http://www.fmrib.ox.ac.uk/fslcourse/lectures/melodic/
> _3200_denoise.fpd/index.html) for more information about denoising.
> I could
> not follow the link, so I looked in the melodic lecture and found
> mention
> of denoising, but I am still not sure exactly how to accomplish
> this. Is
> there a paper that has more detailed methodology?
>
> Lastly, we have a small debate going on in our lab and I was
> wondering if
> you could weigh in. Is it appropriate to use ICA to find components
> when a
> task paradigm is used? Will there by any bias introduced by the
> fact that
> much of the signal and variance are driven by the task, such that
> non-task
> related activity might be more difficult to identify?
>
> Thank you for any insights you can offer:~)
>
> Christina
____
Christian F. Beckmann
University Research Lecturer
Oxford University Centre for Functional MRI of the Brain (FMRIB)
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
[log in to unmask] http://www.fmrib.ox.ac.uk/~beckmann
tel: +44 1865 222551 fax: +44 1865 222717
|