I am designing a multi-center fixel-based analysis study. We have a practical challenge:
Data from Site A was acquired with a single-shell protocol (b=2000 s/mm²).
Data from Site B was acquired with a multi-shell protocol.
Following standard pipelines, Site A’s data would be processed with single-tissue CSD, producing the wmfod.mif file. Site B’s data would be processed with multi-tissue CSD, also producing a file called wmfod.mif.
My core question is: Can these two wmfod.mif files — one from single-tissue and one from multi-tissue CSD — be pooled together as direct input to a single run of population_template?
My technical concern is that despite the same filename, these FODs are mathematically different. The single-tissue FOD is known to contain spurious grey matter/CSF signal, while the multi-tissue FOD represents a “purified” white matter density.
Would mixing them:
Cause the template construction to fail or produce a biased average?
Mislead the internal registration steps due to the different signal scales and contaminations?
What is the correct and most robust approach in MRtrix3 to handle this mixed-data scenario for a unified analysis? Is downgrading the multi-shell data to single-shell (and using single-tissue for all) the only reliable way?
Understanding this design philosophy would greatly help in planning studies with heterogeneous data.
I would definitely not recommend this, as the contrast between the two subgroups will consistently differ (in particular at the GM/WM interface) and the registration would almost certainly suffer from this greatly. Also, the final template would contain a weird mixture of the contrasts from both configurations.
Mixing results from multi-shell data and single-shell data is a bad idea. Single- vs multi-shell is not just a technical difference, but it inherently changes the meaning of what you are measuring. So I would avoid that at all cost.
If you must use all data in a single study, I think downgrading the multi-shell data to single-shell data is probably the only way to go. However, this will not mean that you are now in the clear at all.
Even if both groups are single-shell, they are still coming from different sites with potentially different scanners and different protocols. You would almost certainly have to use per-site average response functions for the deconvolution step, in order to already remove as much of the site-specific response differences as possible.
And when performing the statistical tests you would have to take into account both the differences in variance for each site and the possibility of a systematic deviation between the measurements from each site (“scanner effect”).
It is also important to realize that, it would be impossible to detect statistically significant differences between a group of subjects that has been (almost) completely scanned on site A and a group that has been (almost) completely scanned on site B, because any biological difference between both groups could as well be explained by the difference in site.