Dear @rsmith and the MRtrix team,
We are performing a graph-theory analysis on structural connectomes, and as part of the SIFT2 preprocessing steps, we extracted the SIFT proportionality coefficient mu.
When plotting the mu coefficients of our sample, we detected that some coefficients may be considered outliers (>3SD from mean).
Should we treat mu coefficients like any other dependent variable, and assume that outliers are not representative of the population (possibly excluding these participants), or are outliers in mu should be treated differently? And in general, are there any reasons that make extreme mu values more common than predicted by chance? e.g., measurement noise, improper preprocessing, other biological variables like brain volume, etc. (in the case that extreme mu might indicate acquisition or preprocessing issues, can you give us some clues? We might like to investigate that further).
Other relevant information about our preprocessing pipeline: we used a group-average response function for pre-processing, fixed number of streamlines and dynamic seeding.
Thanks ahead for your assistance,
Typically, three factors contribute to the proportionality coefficient:
- Total streamlines density (as sum of lengths, not just count)
- Total AFD
- The processing “mask” (not actually a mask) defining the extent to which each voxel contributes to the model.
If you have population outliers in mu, then they will be population outliers in either 1 or 2.
If you are generating a fixed number of accepted streamlines per participant, then 1 can only vary due to a change in the streamline length distribution. Maybe some subjects get more or less short association fibres reconstructed as there’s a propensity of them just above or below the minimum length criterion (I’ve been recently debating drastically shortening the default minimum streamline length under ACT). Maybe some subjects have more or less U-fibers, or they are easier or harder to reconstruct. You’d want to see the full streamline length distributions for typical vs outlier cases.
We have seen in some patient cohorts that there can be essentially a global decrease in fibre packing density (/ AFD / intra-cellular volume fraction) throughout the entire WM. If using a common response function, this would lead to reduced WM ODF magnitude everywhere. The other factor I worry about here is global differences in WM T2: if a subject has CSF / GM T2 relatively consistent with a cohort, but shorter WM T2, this will be misinterpreted as a global AFD decrease (and
mtnormalise intentionally won’t fully correct it under typical usage).
Ideally brain volume shouldn’t correlate with mu that much. For a larger brain, there will be more voxels in the model, likely more total AFD assuming that the WM is proportionally increased in volume with the rest of the brain, but also the streamlines will be longer in aggregate.
Hope that helps to isolate