How about defining "ROIs" based on the melodic_IC file (you can use the p = 0.50 mixture model proportions), then apply these as ROIs as masks over the stage-2 maps? The values within each ROI and subject can then be summarised using the average across voxels or perhaps even better, the first principal component. This in turn can be used in follow up analysis. This approach was used in this paper [disclaimer: I was a co-author].

In FSL you can obtain these averages or principal components using the command "fslmeants".

Hope this helps.

All the best,

Anderson