Dear Rui,

one of the strengths of the ICA method (masked or unmasked) is to separate true signals from noise. However, this holds true only if the temporal and/or spatial structure of the noise is sufficiently different from the signal. So under normal circumstances (m)ICA does a reasonable job at separating signal from noise. In that case you will end up with a set of components (ICs) some of which will be signal and some will be noise. The question how to identify the signal components is an active field of research. The most advanced approaches to this are FSL's ICA-based X-noiseifier (FIX), ICA-AROMA and others. However, they are designed for whole-brain ICAs. In masked ICA I recommend the approach laid out in Beissner et al. (2014) - Advances in functional magnetic resonance imaging of the human brainstem. Neuroimage. It takes the local components and checks their global (i.e. whole-brain) connectivity. If this connectivity is not primarily in grey matter, it is more likely for the components to be noise. You can do this with the toolbox by using the back-projection function to whole-brain unsmasked data and analyze the resulting copes or zstats with tissue probability maps of grey matter, white matter and CSF (see my paper for details).

As for your second question: I think most fMRI people would use their entire dataset to estimate the set of components (an in our case: the dimensionality). This approach should give you the most precise estimation of the ICs.

Hope that helps,

Florian


Originally posted by zhao rui: