Upsampled Images for Steps 4 & 5 of Fixel Analysis

Dear Experts,

Should the upsampled FOD and mask images be used in steps 4 and 5 of the fixel analysis to create the FOD population template and subsequent individual subject registrations? The documentation does not explicitly state to use the upsampled, however, the following steps do. Also, if the upsampled images were used in steps 4 and 5 would this majorly be okay or will it detrimentally affect the outcome of the analysis?



Hi James,
Yes, we recommend population template creation and image registration should be performed on up-sampled images. In my experience non-linear image registration benefits from using up-sampled images (the FOD template is sharper with less blurring of neighbouring fibre pathways, than if you create the FOD template, then up-sample for comparison).

That said, you don’t have to up-sample. You are likely to get similar results on standard resolution data, unless your group effect is likely to involve subtle morphological changes.

Note this is just my experience, and I have no concrete data to back this up.

Perfect! Thank you Dave. Much appreciated.


A few reasons why this makes sense:

  • Just for “simple” registration, the benefit is small (but exists): you can do a more advanced form of interpolation (e.g. cubic instead of linear) up to a desired resolution beforehand, and the gradients will also be based on this (potentially following the curves of the edges closer).
  • For template building (the process as well as the outcome), there’s an additional benefit: all involved images have sampled the anatomy at different grid locations (relative to the local anatomy in any given region or structure), so you effectively have higher resolution information on those structures; up to the extent that they can be topologically matched of course. Think of a structure like the anterior commissure: at a typical acquisition resolution of (2~2.5mm)^3, it’ll sit almost in the middle of a few voxels for some subjects, but may seem to cover 2 voxels (at a lower intensity) in others; all due to different alignment of the voxel grid relative to this structure. Combining that information across subjects (or acquisitions) will benefit to define it. The remaining degree of blurriness still present in a template is then due to uncertainty of matching the structures spatially and/or the natural variance in topology (the latter more prominently closer to the cortex).
  • Because the template can then be higher effective resolution in some regions, individual registrations of subjects to the template benefit in their own right.

Thank you for the additional information. This makes things much more clear. I appreciate it.


1 Like