Print

Print


Dear all,

I did group and single subject ICA's (both fixed on 20 ICs) on resting-state data (rats!). Apart from motion correction, no further "cleaning" was done during preprocessing (only normalization to template brain and spatial smoothing). For comparisons, i want to do seed-roi analyses, and therefore i want to do additional "cleaning" of the data, by regression of noise components from the ICA analyses (using fsl_regfilt).

In the single subject ICA's, it is sometimes quite hard to label the components strictly as signal or as noise. The components of the group ICA are considerably easier to divide into signal or noise components. My question is, is it viable to use these group components (actually the single-subject versions of these group components after dual regression) to clean the data by regressing (the single subject versions of) the group noise components from the single subject data before doing the seed-roi analysis?

Of course, i don't get rid of noise which is only present in one or two animals, because this won't get visible as a group component. But there are some Noise components which seem to be present in most animals (e.g. sagittal sinus component, edge artifacts in the first/last slice, susceptibility artifacts near ear canals etc.), which are easy to spot in the group ICA, but sometimes difficult to see in the single subject ICA's (eg. due to splitting of these noise components in several small noise ICs, or due to inclusion of brain parts in these noise ICs which actually look more as if the might be signal, i.e. no clear noise components but rather mixtures of noise and signal).

I actually did this with some animals, and did new single subject ICAs afterwards (1). I also did the cleaning using the noise ICs from the single subject ICA's, and also did new single subject ICAs afterwards (2).
When i compare these two approaches, i see two things:
Using the second approach, the signal components which were either not there/not identifiable or which were at a late position (i.e. IC 15-20 from 20 ICs) move to a relatively early position (i.e. IC 1-5 from 20 ICs) in the second ICA after cleaning.
Using the first approach, this effect is not as strong (i.e. a signal component only moves from IC ~17 to IC ~10 or so, instead of IC ~5), but the spatial appearance of the ICs actually looks better (i.e. they look more like the equivalent signal ICs from the group analysis). And this approach is much easier (as no labeling of the single subject ICAs needs to be done).

So, my question again: is this approach (using the single subject versions (after dual regression) of the noise ICs from the group ICA) viable? Or should I stay with the single subject ICAs for cleaning the data?
I am a little afraid I might introduce things that are not there using this approach... comparable to the story with the whole-brain-signal regression and the introduction of negative/anticorrelations of brain regions.
Probably one can first regress the group noise ICs, then do a single subject ICA, and then regress the remaining noise ICs from this single subject ICA, and then do the seed-roi analysis with this data?

I would welcome your opinions.
Greetings,
Andreas

########################################################################

To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1