Thanks for the comments. It helps a lot to figure out what to do. However, I have another questions for your comments
1. "that's largely empirical I think - within reason it could make sense, though it's not very obvious what the practical pros/cons would be."
>>Actually the reason why I restrict the component area by giving the threshold is that I would like to compare the time-courses estimated within the group ICA mask between participants like the case that fluctuation differences between before and after the a certain medication (for example). So I thought that if I didn't restrict spatial-map boundary to a certain robust region of a given component during the stage 1, then I would have too global time courses from a kind of too-much-overlapped brain regions across all component maps. Actually when I indeed tried to use thresholded one to estimate time-course, the results are pretty similar with unthresholded one in terms of correlation coefficient.
2. "you can extract regional summary measures from a given dual-reg spatial map output - but this *only* tells you about "connectivity to" this RSN - not all others - so it is not a general measure of connectivity from a given region to the rest of the brain in the context of the full set of resting-state networks."
>>In that case, i wonder it is okay to feed individual estimated time-course (stage 1) into seed-based correlation analysis (e.g., FEAT), and then extract mean signal from a given ROI (e.g., amygdala). That is, putting individual time-course of DMN in the GLM model to compute correlation values over the brain region, and calculate mean values with my ROI mask.