One the main doc, I see:
fMRIPrep performs minimal preprocessing. Here we define ‘minimal preprocessing’ as motion correction, field unwarping, normalization, bias field correction, and brain extraction
The above statement makes it sound like it performs bias correction of the BOLD time series, but from what I can see in the workflows and outputs, it seems to only estimate and apply a bias field for the anatomical images. Is this correct? If so, is there a recommended approach for handling the bias field in BOLD data?
Bias field correction is done on the bold reference to improve coregistration, but it is not applied to each bold volume.
As far as I know there’s not a recommended approach to handling bias fields in BOLD. https://www.sciencedirect.com/science/article/pii/S1053811912010804 argues for correcting before motion correction; practiCal fMRI: the nuts & bolts: Review: Using a bias field map to improve motion correction of EPI time series argues for trying both and evaluating the results before committing.
The way to do this with fMRIPrep would be to bias-field correct volumes before passing to fMRIPrep.
Would it make sense to incorporate this, or at least provide estimated bias field maps? The HCP pipeline computes a bias field from the SE fieldmaps and applies them at the end of the fMRIVolume pipeline. The same approach could be applied to the magnitude from phasediff fieldmaps.