Hi,
I am trying to do a group ICA using the temporal concatenation approach, with the ultimate goal of running dual regression to compare groups. However, I have two (related) issues regarding the 4D input files.
1) I have multiple runs per subject, but if I'm interested in individual differences, I assume I need one input for each subject. So, I've concatenated the runs for each subject to create one input per subject; however, I'm wondering if there's a "best" way of doing this to account for different scalings in each run and the transitions between runs. I see at least two approaches: (A) concatenate the fully preprocessed data (motion correction, BET, filtering, smoothing, slice timing correction) runs for each subject; or (B) do "enough" of the preprocessing so that each run gets "grand mean" scaled to 10,000 (it seems pre-stats will do this even if nothing else is being done?), concatenate those files, and then run the full battery of preprocessing on the concatenated dataset. With the first approach, there will likely be some sharp transitions between runs, so hopefully the second approach will ameliorate that issue since it's doing motion correction across the whole concatenated data series for each subject.
Unlike FEAT, I don't see a clear parallel where I could simply do a fixed effects analysis across each subject's runs at the 2nd level and then combine those outputs at a 3rd level group model in MELODIC… Am I missing something?
2) The number of runs across subjects isn't equivalent because some runs need to be dropped due to excessive head motion and some runs just had more or less time points. It seems the group-ICA expects each 4D input to have exactly the same number of time points. Is there any to make this flexible? I'd rather not just go with the minimum amount of data across all my subjects, especially if I could just populate the "set fmri(npts)" field with whatever fslnvols returns for each input/subject.
Thanks!
David
|