For the study cohort of stroke patients with obvious lesions and healthy controls, how to generate the template for the FBA analysis of single-tissue CSD? Should a single template be generated for the full study cohort (i.e., both patients and healthy controls), or should a unique template be generated for each group?
(1) If the previous (single template for 2 groups) is better, how to mitigate the influence of white matter lesions on template generation, including warping DWI to template space and DWI metrics?
(2) If the latter (unique template for each group) is better, is it feasible to perform whole-brain statistical analysis (such as CFE, fixelcfestats command, etc) of fixel metrics between 2 groups using 2 unique templates for each group?
(3) To solve this problem, could I merely introduce the stroke lesion volume as a covariate to control the effect of stroke lesion if I use a single template for 2 groups?
P.S.: Here is a published FBA article of stroke patients (Natalia Egorova et al, Pervasive White Matter Fiber Degeneration in Ischemic Stroke, Stroke. 2020;51:00-00. DOI: 10.1161/STROKEAHA.119.028143). “A population template was generated using fiber orientation distribution images from 30 (15 stroke and 15 control) participants”.
That is to say, this study uses a single template generated for both patients and healthy controls. Is the method appropriate to generate a population template containing stroke patients? Will it be argued by reviewers (but this article using a single template for stroke patients and healthy controls has been already published in the famous journal Stroke)?
It’s difficult to know for certain exactly which approach is ideal in such scenarios, and it probably depends to some extent on the magnitude of the pathology. To understand the pros and cons of different approaches here requires appreciating the rationale behind the relevant building blocks of the pipeline.
Firstly, this question needs to be separated out into “template generation” and “DWI metrics”. You presumably want to be computing DWI metrics everywhere regardless of lesion presence; the question is how the presence of such lesions may influence template generation.
This can then be further separated out into “do I include stroke subjects in the generation of the template?” and “If yes, how do I mitigate the influence of white matter lesions?”.
The former is the primary decision to be made in such an experiment: do you use healthy controls only to construct a template, or do you also include stroke patients? Recall that the benefit of using a cohort-based template rather than a fixed template is to minimise the “distance” between your cohort subjects and the template, which makes registration more robust. Now imagine that you have an exemplar stroke subject with extreme pathology. By including that subject in template generation, the resulting template is going to look a little bit more like that extreme pathology. That has the prospect of improving the registration of that one subject to the template, but potentially at the expense of the registration of all other subjects to the template. Conversely, if that subject is not used in generation of the template and you are doing explicit registration of that subject to the template, there may be some risk of the registration going completely awry; but at least that error is constrained to just that one subject, rather than contributing such corruption to the template itself. So it’s a balance of different types of errors.
In general, I’d suggest trusting your intuition. If you look at the stroke images and think “if I include this in my template generation it’s going to ruin my template”, you might want to think about omitting them from the template generation, and then make sure to check manually the subsequent registration to the template to determine whether or not that subject needs to be excluded from the study.
For the second part—what to do about lesions—there should be the opportunity to mask out lesions both during explicit mrregister calls, and if such masking needs to be performed within population_template. Some of this has changed with software version 3.0.0; @maxpietsch is the expert there. If you read the help description for the -mask_dir and -nanmask options of population_template, that explain how you can both prevent voxels from driving the registration process, and prevent them from contributing to the resulting template; I suspect in your case you probably only want the former.
If the two groups have been registered to two independent templates, there’s no way to perform statistical inference. The software needs to be able to look at an individual fixel, aggregate the data within that fixel across all subjects, fit the GLM and extract the measure corresponding to the group difference; if the data are in two different fixel templates, there’s no way to do such.
A generous interpretation of the question would be that you generate two independent templates—one for each group—and then do a second template generation step that aggregates the two templates together. This would however be a bad idea. If a statistically significant effect were observed, the confidence with which you could attribute such to a genuine group difference, as opposed to imperfect registration between the two per-group templates, would be very low.
I would consider the inclusion or exclusion of lesion volume as a nuisance regressor to be a question wholly independent of template construction & registration; it doesn’t “solve” any problem in that context.
Whether or not to include such in the model shouldn’t be seen as a matter of preference; it fundamentally changes the scientific hypothesis being tested. In the absence of such, your observation of a statistically significant effect would result in the statement e.g “there is reduced FD in this bundle in stroke patients compared to controls”; in its presence, you would say “there is reduced FD in this bundle in stroke patients compared to controls, over and above that which can be explained by total lesion volume”.