Strange artifact in DWI - differential slice-wise scaling in AP directions

preprocessing

#1

Dear experts,

I am writing this post in the hope that you can help me to correct for following artifact I encounter in the data of one of our studies. This artifact is manifesting by differential slice-wise scaling in AP direction (AP direction is phase-encoding direction). The scheme is 2 slices stretched followed by 2 slices squeezed an so on. The artifact manifests itself only in some specific diffusion sensitization directions (in about 10 from 64 directions) and AFAICS it seems that it is present in all of the subjects in the study (consisting of approx 200 subjects). Some of directions are not affected in all subjects, but there are several specific directions which are affected in all subjects.
It is standard Siemens Trio single-band 64 direction sequence, b=1100. Standard interleaved acquisition (that is why it is puzzling me since in the artifact there is interleave in slice scaling in period of 2 slices, not in one slice).
See the screenshot demonstrating the effect:
DWI vol with artifact:


subjequent DWI vol without artifact:

I hoped that the new slice-to-vol registration in eddy could be able to correct this artifact (Jesper expressed hope that it could be the case in the personal communication) but according to my tests with the new eddy it is not the case, the artifact is present even in the eddy-corrected data.
It is probably not the movement problem since the slices are not moved to each other, there is systematic scale between them.
Do you have any idea of the source of the artifact and how to correct for that?
I think it could be doable to correct it since the artifact is so systematic.
I was wondering whether I could do following (probably not ideal approach, but it is better than nothing):

  1. split affected volume into two separate vols with stretched and squeezed slices
  2. register two vols together by restricting to movement and scale to AP direction (it is possible to do it via mrregister?)
  3. concatenate mutually correctly scaled volumes again together

Do you have better idea?
Any comments/suggestions will be greatly appreciated.
Regards,

Antonin


#2

Wow, I’ve never seen anything like this… If this is a product sequence, and it’s running in traditional single-band mode, I would definitely chase it up with Siemens - it might indicate an issue with your scanner, or a bug in the sequence code that they’ll want to fix ASAP…

Is there any consistency to the directions affected? For instance do they tend to cluster around the PE direction? But even then, it’s hard to envisage any hardware-related issue that could cause this, it looks like a bug in the acquisition or reconstruction code…

As to how to fix this, your suggestions sound like a good starting point, I’m not sure I can add much. The one thing I’d change is maybe trying to avoid registration altogether, and see whether the scale factor that you need to apply is actually reproducible across slices and subjects - in which case you might be able to use mrtransform with a single affine matrix, and avoid errors introduced by registration (you might need to remove or reset the image’s transform matrix for that work if you have oblique acquisitions - and then put it back in afterwards). I’d personally be concerned about having up to scale up the affected slices, it’s not unlikely that the required interpolation will introduce some systematic bias… If the number of volumes affected is not too large, I’d be tempted to discard them. But that might be problematic if the affected DW directions are all similar, since that will likely introduce an orientation bias… Not a simple problem to fix!