Limits on rotations when combining EPIs?

[effigies]

Inspired while thinking about #297: for SyN-SDC in fMRIPrep, we collate BOLD series based on their effective echo spacing/total readout time and phase-encoding direction. If you consider cases where subjects adjust their heads between runs or leave the scanner, it seems like there is potential to have significantly different distortion profiles.

Consider a case where in session 1, the subject's head is rotated ~15deg to the left of plumb about the vertical axis (in the axial plane), and ~15deg to the right in session 2. You would now have a 30deg difference in the direction of susceptibility artifacts between the two scans. An average EPI reference would increase noise, rather than reduce it. Rotations about the transverse axis will also impact the distortion.

If this reasoning is correct, it's possible that this is responsible for some SyN-SDC artifacts in multi-run datasets. I will be testing a multi-run dataset on a per-run basis (via the expedient of just changing the TRT because it doesn't matter).

If this is the case, perhaps the "correct" solution is to estimate the fieldmap in T1 space and average that across BOLD images, rather than averaging the BOLD references and estimating the fieldmap. A compromise for efficiency could be to estimate the alignment between BOLD references and cluster images that are sufficiently close together.

[effigies]

Per meeting, try to get wrangler to generate 1 fieldmap per BOLD. If this works, get it into 2.2.0