Hi @Alistair_Perry,
Essentially, @jdtournier (see quote) is right here, and there’s no immediate reason for concern. However, there’s a few potential additional reasons why this may arise from the current algorithms, especially the mechanism many of them share to select the single-fibre white matter response function. Still not a big concern, but there’s some room for improvement indeed (although I doubt that would ultimately be the explanation/“solution” to the non-decreasing coefficient magnitude you observe; that in and of itself is probably not even a problem to begin with). I’m looking into this currently, with a few solutions in place even, although their robustness remains to be tested at my end. I’m not one to sacrifice robustness over some accuracy in all cases… so well, I’ll see what the testing gives. @Alistair_Perry, if you’re really concerned for your particular application or project, happily be in touch. I’m already trialling some changes in a few projects with external collaborators; but of course safely supervised at this stage.
Well, certainly not commented by me, that would be.
Oh yes, of course. The algorithm leverages as much of the data as it can, all within Occam’s razor’s principles of course (no need to try and get from the data what isn’t in there). All clearly documented in the original abstract. Also, if people are still wondering, here’s some additional insights on how msmt_5tt and dhollander perform with respect to each other. So far, that pattern has generalised over all data I’ve encountered so far, including ranges of lesions, more lesions (also, we’ve used it in MS studies, and it’s been used internally on a stroke cohort, and externally on a few other pathologies), animal data, ex-vivo data as well as babies. Someone at our recent workshop had single-shell b=800 data with a low number of directions (don’t recall the exact number), with a massive cyst in the brain, and an equally massive MR artefact that distorted part of the brain and basically destroyed the signal in that area… well, dwi2response dhollander still selected sensible voxels. But the message here should always be: check your -voxels output. If it looks sensible, you’re good.
But well, your results will change depending on which b-values you feed into it; that’s entirely expected and by design… no need to worry about that.