Population template for large number of subjects

[Romain_Valabregue]

Hello

thanks again for this very nice software.

Running population template for many subject takes a lot of time. After some point adding more subjects should not change the template … no ?
So if one have a big cohort (N>100), may be an option is to perform a population template on a subset of subject (N=10 , 50 ?) , and the just do a coregistration of the remaing subject to the template.

but I recently come up with this comment from @maxpietsch

@maxpietsch

If you don’t have warps, you can re-register the subjects’ FODs to the template (as in the FBA pipeline) and use these warps to transform the subjects’ FA images to template space but the two templates wouldn’t be 100% comparable as registration procedures inside population_template differ from directly registering a subject’s FOD to the final template.

How much much different. ?
Will it be possible to get something similar to the population template (as if we were in the final itteration …) ?

[maxpietsch]

@Romain_Valabregue

Will it be possible to get something similar to the population template (as if we were in the final itteration …) ?

population_template iteratively builds the template using increasing resolution and degrees of freedom and refines the transformations in each iteration using the current estimate of the template which should sharpen over iterations. There is no shortcut to estimating the warp for the second to last iteration and adding subjects would slightly alter the template in each iteration. So it’s not possible to retrospectively generate warps for new subjects in a way that is consistent with the template warps.

However, our recommendation is to (re-)register all subjects directly to the final template (Fibre density and cross-section - Multi-tissue CSD — MRtrix3 3.0 documentation). Assuming the template is representative of the cohort, this results in consistent warps. The compute required scales linearly with the number of subjects for both population_template and the individual registrations to the final template but template creation iterates multiple times over all subjects so is typically much slower.

@Romain_Valabregue

How much much different. ?

That depends.

On a related note, optimal registration parameters to generate a “good” template might be different from generating a “good” mapping from subject to template space for individual subjects. For instance, the desirable degree of local deformations might depend on whether the goal is optimal template sharpness or morphological analysis. There are unfortunately no easy metrics or heuristics to optimise these without inspecting the results.