No worries. The “trickery” I was referring to was actually something slightly differently: TractSeg outputs “tract orientation maps” (they call them “TOMs” as I wrote, I believe). Together with the voxel-masks of the tracts, those TOMs are actually the “real” output from TractSeg itself. They’re basically peak orientations, which indeed makes them very much like fixels! So the trickery I was hinting at would be first making those peaks into actual fixels (this only refers to the data format, nothing really changes to what they are otherwise) via
peaks2fixel. That would then be a “stand-alone” fixel image so as to say, with it’s own
directions.mif (with only fixels for a given tract/bundle). Then I’d make some handy use of
fixelcorrespondence to “map” those onto the fixel analysis mask (the one that you’d have derived earlier in the pipeline from your FOD template).
However, in the meantime TractSeg also outputs streamlines (
.tck) itself indeed. This is not a “direct” TractSeg output, in a strict sense: those streamlines don’t pop out the neural net itself, but they’re computed via the TOMs actually. In short, the TOMs are turned into “FOD-like” ODFs, and those are then fed to
tckgen (all wrapped in the TractSeg code).
But in some ways, the latter is actually a great thing: by applying actual streamline tractography on those ODFs-derived-from-TOMs, the “tract” (streamlines) you get is inherently more regularised, i.e., smoother; whereas the TOMs themselves might be (slightly) more noisy depending on the original data quality. So then using those streamlines with
tck2fixel, as you suggest, will actually be more robust in the end.
In practice, this has now also become much more convenient since the TractSeg package itself directly does all the steps up until and including generating streamlines for you, and it does it very well (in my experience). So I’d say: for most purposes, ignore the the “trickery” I was originally referring to, and go with
tck2fixel; for both practical simplicity as well as improved robustness. We’ve been using this in our lab now routinely to generate fixel ROIs in templates; both for (fixel) ROI-based analyses as well as almost fully automatically labelling whole-brain FBA results themselves, i.e., as a somewhat more objective and standardised way of reporting FBA results. It works extremely well if your data has been processed with any form of 3-tissue CSD (either MSMT-CSD or SS3T-CSD), since that’s the kind / quality of WM FODs TractSeg has effectively been trained on.
Hope that helps; feel free to ask if not clear.