afni.nimh.nih.gov

It is best, as Gang implies, to extract the global signal and then include it in your regressions in the pre-processing, which presumably are removing other signals "of no interest". Regressing one thing out by itself after the other regressions is not usually a proper form of analysis -- if you regress out A and then B, the result is not the same as regressing out B and then A (unless A and B are orthogonal).