We are interested in directly comparing the performance of ME and SE scans for deriving subcortical signals.
If one wanted to apply GSR in a manner that is more comparable to how it is implemented in SE studies (e.g. Redirecting), and if one is only interested in the combined ME image, would it be reasonable to regress out GSR after tedana runs PCA/ICA? And, if that sounds reasonable, should GSR be recomputed from the preprocessed cleaned combined ME image or should one compute GSR from the original ME image?
Also do you have an opinion on the order of applying GSR, spatial smoothing and BPTF?
In typical SE studies, folks regress global signal out as part of a general denoising step, including other confounds (e.g., motion parameters), rather than in a separate step, so I would recommend doing the following:
Extract global signal from the optimally combined data (not denoised).
You could use the gscontrol options within tedana, but the global signal regression method we implement is pretty different from “standard” GSR, and the minimum image regression option is definitely not comparable.
Run tedana with --tedort enabled.
Grab the orthogonalized rejected component time series from the tedana outputs.
Combine those ICA component time series with the global signal in a single array.
Regress out the combined array from the optimally combined data. If you want to do temporal filtering, do it in an omnibus step here.
Do any spatial smoothing you want.
There are plenty of other ways to do it. You could extract global signal from the ME-ICA-denoised data, but that’s not quite as comparable to SE processing since folks don’t generally estimate global signal after regressing out other confounds.
I generally agree with @tsalo. One thing I’ll emphasize is that GSR will almost definitely remove both neural and non-neural signals. IF multi-echo denoising works as intended, and comparatively more non-neural signal is removed, then calculating the global signal after the full denoising means it would contain relatively more neural fluctuations-of-interest that would then be removed with GSR. This is why it’s better to calculate the global signal on the optimally combined data.