Resulting from discussion here.

eddy assumes intrinsic alignment between the first of its b=0 input volumes, and the first volume that was input to topup. There is however a dwipreproc usage scenario where this isn't the case: The -rpe_pair option (or, more generally, using the -se_epi option in 0.3.16). Any mis-alignment between the first volume acquired for susceptibility distortion correction, and the first b=0 volume in the DWIs, will result in incorrect inhomogeneity field correction.

There's two ways I can think of that this could be dealt with (that don't require modifying FSL or superseding those commands), but I want to make sure I pick the right one:

1. Add first DWI b=0 volume to topup input

Take the first b=0 volume in the DWI series that is to be fed to eddy, and concatenate it to the start of the image volumes to be fed to topup. That way, when topup generates all of its results with respect to this volume, it's intrinsically aligned to the volume that eddy subsequently uses as its reference.

• May introduce a bias in topup, e.g. estimating the field based on 2 A>>P and 1 P>>A volumes. Could potentially look for this in the phase encoding table and scrub a volume accordingly.

• I think this should intrinsically account for differences in header transform between the single b=0 and the other SE EPI volumes, if my assumptions about topup and eddy are correct, but I'll need to test.

• If FoV size is not identical, would require re-gridding the b=0 volume to match the SE EPIs (or the other way around).

• Assumes equivalent TE == equivalent contrast between b=0 volume and SE EPIs (which wouldn't be the case for our protocols if I had my way).

2. Explicit registration

Perform explicit rigid body registration step to align first b=0 volume with first SE EPI volume.

• Would require more exhaustive testing to make sure everything's behaving as it should. E.g. My suspicion is that the b-spline coefficients are derived in image space and then applied in image space with no regard for the header transformations, which means modifying that transformation may do nothing.

• Would be able to deal with differing contrasts between the b=0 and SE EPI volumes.

• Registration would be imperfect since the first b=0 volume will still be inhomogeneity distorted when the registration is performed, and it's not guaranteed that an SE EPI volume can be found with equivalent distortion. Solving this would require performing iterative registration between the topup-corrected brain volume and the eddy-corrected volume during eddy.

I think Option 1 is winning out having written it down. But open to comments. Also open to help given I'm supposed to be grant writing 👎