I would like to extract AFD from a particular set of tracts that I have a priori hypotheses about and relate their AFD to behavioural measures. The documentation on the website speaks specifically to the fixel-based analysis and I was wondering if my approach is valid.
I skull stripped the diffusion weighted data and preprocessed the data with fsl topup and eddy. I ran dwi2response on all subjects in the group and then took the group average of the output. I used those group averages to calculate the FOD files for each subject with dwi2fod, Performed joint bias field correction and intensity normalisation and generated whole brain tracts and then pruned them to my desired ROIs, which I would now like to use as masks to extract AFD for each subject and relate those values to behaviour.
Will these values be relatable across subjects or does the analysis absolutely have to take place on a fixel level in template space?
3 shells, 60 directions, multiple b0’s; let me know if any other pertinent info is needed.
Best to break the question into two somewhat separate concepts:
Intensity normalisation: While the methods used have changed slightly from the original 2012 AFD manuscript, the concept still holds: You need to perform explicit steps in order to make AFD comparable across subjects. You’ve used a group average response function and applied the multi-tissue bias field correction / intensity normalisation method, so everything is covered there based on our current recommendations.
Granularity of hypothesis testing: Performing segmentation of particular structures, extracting e.g. mean values of quantitative metrics within these structures, and comparing these values between subjects, is perfectly reasonable. This is akin to VBA evolving from voxel-wise correction to cluster-based stats: You potentially gain statistical power, but lose the specificity of exactly where inside the “cluster” the effect may be greater or smaller. And of course you won’t detect any differences that are outside of the set of segmented structures that you choose to test.
But a couple of things to be wary of:
- Getting a voxel mask of a particular bundle is not too difficult; getting a fixel mask is a little trickier.
- A whole-brain FBA intrinsically gives (weak) family-wise error control across the fixel template / mask. Extracting quantitative values for individual bundles, and then handling those values externally, can easily turn into an uncontrolled statistical fishing expedition / p-hackathon if you’re not careful.