I have a doubt regarding the use of an average response function for a group analysis. You need to use it for certain analysis, like fixel based analysis or SIFT2-weighted connectome comparison, but what about SIFT connectomes? In this case the meaning of the connectome would be number of streamlines, so I don’t think an average response is needed, am I right? As far as the bias field inhomogeneities over the image are minimized, it should be fine, right? Thanks in advance.
EDIT: My train of thoughts came after reading this discussion, as my idea is to do something similar, a pipeline to provide multi-contrast connectomes. Then I was wondering if would be more accurate to just run tcksift and then provide all the connectomes with the final set of tracts (without providing the FBC), rather than using tcksift2 without an average response function.
If you can follow along with the logic for inter-subject connection density normalisation for SIFT2 as per the article, then I would encourage considering the following quote:
(Note that in the original SIFT method, ws = 1 for all retained streamlines after filtering, but µi is modulated during the filtering process.)
SIFT uses the same model as does SIFT2. It just can’t provide a continuous weight per streamline; it’s a one or a zero. And mu is different after filtering. The software may not (currently) yield the model fit in this form, choosing to export a filtered tractogram file rather than a vector of (binary) weights, but the interpretation of FBC holds.
Reconstructing a fixed number of streamlines for each individual is a way to do inter-subject connection density normalisation; but not the best way. I could have expanded Figure 13 in that manuscript to include a proposal where a fixed number of streamlines was reconstructed for every individual. In that scenario, every single one of the 16 cases would have exactly the same “connection density” despite obvious underlying differences.
Another point to consider: If you were to do a mass univariate statistical comparison of number-of-streamline connectomes between two groups, where you have reconstructed an equal number of streamlines for each participant, then observed differences are guaranteed to manifest as both increases and decreases in equal proportion. If you’re not comparing per-edge connection densities and instead looking at some other derivation, then whether or not these steps are consequential depends on whether that analysis is invariant to global scaling of the connectome.
It’s clear that I need a second manuscript with more discussion on the inter-subject connection density normalisation issue; though I’ve no idea when I’ll ever find the time to get to it… I did start writing a new documentation page to accompany these changes that will explain things from a slightly less conceptual and more pragmatic perspective along with the interface change foe 3.1.0, I’ll try to add some info about SIFT1 in there.