Estimate response function for multi-shell ex-vivo marmoset data

It’s late-to-the-party-day, but here we go anyway :grin: :

So well, nope, not really a problem after all. :man_shrugging: The basic strategy and relative tissue properties it relies upon still hold for a marmoset. The challenge with these kinds of data is probably rather the ex-vivo aspect versus the maximum b-value you’ve been able to acquire. This indeed:

I’ve heard similar numbers, varying from a factor 3 to 4 indeed. But it also indeed seems that this was at least a bit anticipated here as well. If this lands you with an “in-vivo equivalent” b-value of about 2000, you should be good. What may be needed though, is to tune the -fa parameter of dwi2response dhollander, as I advised here as well: Multi-tissue CSD . A rodent (though in-vivo) is quite a step further away from a human-like brain compared to a marmoset… and even there, no worries given an appropriate tuning of -fa.

Sometimes not a bad idea indeed, although I’ve found it’s actually surprisingly hard to really select good voxels manually. Even single-fibre ones: you may aim for traditional single-fibre regions, only to select voxels that actually have quite some dispersion (though no “crossing”). The tournier algorithm avoids these consistently (and rightly so).

Too bad… :slightly_frowning_face: But this was crucial to dwi2response dhollander working correctly though as well: an initial FA threshold (i.e. that -fa option) is quite an important initial step. If that’s messed up by a faulty FA map, the algorithm may be set up for failure indeed. On the other hand, it’s a pity you now had to exclude this shell, because the highest b-value is definitely the most valuable one for most, if not all, CSD variants.

The result looks great though! The reason it works so well compared to traditional segmentation methods is that multi-tissue CSD is actually not a segmentation method (with e.g. spatial priors and based of just 1 or a few intensities): the signal within each voxel is pulled apart into tissue signal contributions independently. So it doesn’t matter how thin any structure is (spatially): if the tissue contribution to the signal in those voxels is there, it should be able to tease it out.

Makes sense indeed, but may depend on what the ex-vivo brain was suspended in, or even how it was prepared. But I have to agree, judging from the depicted results, it looks ok even if it would just be modelled with WM and GM alone.