Hi David,
> We are interested in delineating network connectivity in our task-based fMRI event related design.
> Basically we have 2 contrasts of interest, A & B. We have isolated 30 independent components -ICs-
> using Melodic and now we want to use fsl_glm to find components with a temporal course
> that is correlated with the temporal course of A and with that of B. Is this a valid approach to infer
> connectivity networks correlated with our task protocol?
Yes, though if the 3 different runs have identical timings then you could also use the tensor ICA approach.
> The first issue is that our fMRI design contains 3 runs per subject. In doing fsl_glm we have to specify the
> .mat and .con files from the FEAT design in order run it. Then we search for correlations between each IC temporal
> course with that of each contrast, by testing the ouput Z score from the fsl_glm against zero (using a t test, correcting the p
> value for the number of components tested).
Correct
>
> Since we have 3 runs we will have to ru fsl_glm 3 times per subject, one for each .mat. and ,con from the 3
> lower level analyses. The question is: could we just average the 3 Z scores we get from each fsl_glm -1 per run- and test that against zero? [in order to test for statistical significance as above] I hope this make sense.
>
Yep, that's one way of doing it.
> This way we expect to derive IC-based connectivity maps that are correlated with our task effects of interest A & B.
>
> My second question is related to the dual regression approach, which to be honest am still struggling to learn whether or not
> it may gives something else than Melodic and fsl_glm. I think dual regression may be useful to derive connectivity maps for each individual, based on the ICs derived at the group level
>
That's right - and with this output you can now do a voxel-wise analysis per component (rather than an analysis which is global per component)
> I wonder whether there may be some utility of the dual regression approach in the context of the analyses I have
> described above.
>
> Finally and I dont mean to pester, is there any short-term plan of getting a dual regression practical in the FSL website
>
Probably not, as the next version of melodic will deal with things quite differently.
hth
Christian
> Thanks a lot
>
> David
|