Hello FMRIB et al:
First, thanks FMRIB for such a great course in Brisbane week before last--
I learned a lot that will definitely enhance my use of FSL and Freesurfer.
Plus, it was fun to meet you in person-- you're all much younger than I
would've guessed! (which is not meant to suggest you sound stodgy and old
via email... just wise!)
Second, as a result of the course, I've been exploring using MELODIC at
the single-subject level for the purpose of denoising my data prior to
FEAT. Many of the components are obvious noise, and I have little worries
about getting rid of them. Others, however, are less clearcut, and so I
was hoping for some input, if anyone is willing. I made a MELODIC report
available here: http://www.tufts.edu/~hurry01/ebbl/report/00index.html.
The components I'm specifically wondering about are 1, 5, 6, 7, 11, 13,
20, and 30. Components 6, 20 and 30 (and maybe 11) look to me perhaps
like aliased cardiac signals, but I'm not sure. All of these questionable
components have very fast signals with a similar power spectrum, with
peaks at about .091146 and .109375 Hz. I would lean towards removing
them, but maybe that would be bad?!
The design I'm using is an event-related design whose fundamental
frequency is .046875 Hz, which is nicely apparent in component 2. In that
component's power spectrum, you can also see a small peak at .09375 Hz,
not surprising since in a sense there are at least two "events" per trial
(e.g., picture appears, instruction provided 4 seconds later, then on to
making a rating), and you can see in component 2 how many of the responses
have little double peaks. In any case, to the extent that more than one
BOLD response is generated in each trial, the task frequency could be
double or even triple .046875 Hz. Clearly I don't want to remove
components that contain purely task-related variation...
I'd love some input from those of you who are accustomed to judging these
pICA components for denoising. Thanks in advance!
Cheers,
Heather
P.S. If there is in fact reason to believe that I have double-responses
in my signal, what would be the best way to model that at the first
level? I'm currently using the default flobs to model the canonical HRF
with a temporal derivative and one that gets at dispersion as well. I
suspect this won't capture the variation of interest... Perhaps I should
generate a double-humped canonical HRF?
|