me and my colleagues are setting up a longitudinal study. We want to create an unbiased template for each subject in the study. The goal is to measure FA changes in some VOIs (not pixelwise).
In your opinion, is it better:
to fit DWI, then create a DTI template
to create a DWI template, using the population_template script and then to fit to DTI?
According to this article, the 2nd option is better in pixelwise approach; we could not find anything on the VOI-based approach.
I have no experience with this but I guess there is no data/ question independent answer as to which approach is better.
I would create an FOD template and use the resulting warps to transform the native space subject FA maps. Although you’re looking at scalar values, have a look at the fixel based analysis pipeline.
Out of curiosity: If a method is better in a per pixel level, then why do you expect it to perform worse on a VOI level?
If you are doing a very simple study where you plan to manually draw the volume/region-of-interest (ROI) on each subject, then you can probably get away with a rigid alignment of the FA maps between time points.
However, if you hare planning on registering each intra-subject template to a group average inter-subject template, so that you can draw a single ROI on the group-average (and inverse transform this back to each subject), then I agree with @maxpietsch that it would be better to register with FOD images, then use the these warps with mrtransform to transform the FA maps or the template ROI if desired.
As a side note, there is no way to currently register the DWI images (or some Q-space representation of them) non-linearly with population_template.
For creating an intra-subject template with population_template, I’d recommend using the -rigid option . Also output the warps with the -warp_dir option. You can then use mrtransform to apply these warps to the FA maps.
One thing I have not yet explored are good default settings for intra-subject non-linear registration. For most studies the difference in brain morphology between time points should be very minimal, so you can probably get away with fewer iterations (-nl_niter option), a multi-resolution scale starting at 1.0 (-nl_scale option), and starting with FODs at lmax 4 (-nl_lmax option)
Final note, If you have more than 2 time points you will need to use population_template. Otherwise if you only have 2 time points you could use mrregister directly, and use the -transform_midway option to get both images into a midway space. You can then generate the intra-subject template by averaging these midway images using mrmath sum.
Out of curiosity: If a method is better in a per pixel level, then why do you expect it to perform worse on a VOI level?
To be honest, I did not expect the 2nd method to perform worse, but not so much better than the 1st. At VOI level errors could compensate each other.
The 1st one was simpler, quicker and there is no need to update bvecs after the transforms.
I think I would follow your suggestion and create a DWI template. Thank you for your help!
I will follow the first part for sure. As concerns the second [about using mrtransform], wouldn’t it be better instead to fit the warped DWIs (of course after having updated the bvecs files)?