Susceptibility correction old DWI

Dear experts, I hope you are doing well.

I have an issue with susceptibility correction in old DWIs
Besides our most recent DWI image protocol, we have hundreds of an old DWI which is a single shell (b-1000), with 32 gradients directions and 1 b0 image.
Since we don’t have an additional image with reversed-phase encoding direction, I understand that the only way to improve the frontal lobe susceptibility distortions is non-Linear registration (right?).
What I am doing is: fsl Flirt registering my T1WI to the DWI space, using a pós processed b0-image (I always prefer to keep DWIs in their original space). Then, using fsl FNIRT I non-linearly register the same b0-image to the previously registered T1WI (now in the same space).
The whole thing runs smoothly and the overlapping is fine, but this non-linear registration doesn’t fix the susceptibility distortions in the final non-linear-reg b0 image, at all. Any help or suggestion of a different pipeline?
Thank you very much.

Dear all,
I believe I understood why I got no answers: MRTRIX manual does not recommend non-linear registration to correct susceptibility distortions. The reasons mentioned are ineffective distortion correction (as I noticed before) plus (much more important) unpredictable brain moves that may affect the diffusion estimation. I will skip the non-linear registration due to the second argument.
Just to register here, I got fine distortions correction inverting the masked b0 image contrast (newb0=abs(oldb0-max(b0)) and registering to the T1 using ANTs (rigid + Affine + sync), but, doesn’t matter anymore.


There is a new method to do susceptibility correction for old DWI using only the T1w. Please take a look here. You can generate the top up field maps and the using it in dwifslpreproc.

Best regards,


1 Like

Hi Brunno,

The concerns around use of generic non-linear registration in this context IMO are moreso:

  • Using off-the-shelf registration procedures and failing to constrain non-linear deformations to the phase encoding axis only (and to a lesser extent to update the rigid-body component upon updates to the non-linear mapping);

  • The ill-posedness of non-linear registration of images that contain different contrasts (it’s difficult enough to get right for images with the same contrast).

I can’t comment on the full gamut of solutions out there, and there will be various discussions on the topic dating back a decade or more; but Synb0-DISCO is tailored specifically for this purpose.


1 Like

Thank you.
I agree with your arguments. I am reading about the Synb0-DISCO method right now.