Threshold option for ss3t_csd

[JustinGarber]

Hi

Is there an equivalent option available for ss3t_csd that the -threshold option plays for the MrTrix CSD script dwi2fod msmt_csd?

I want to play around with the -threshold option to model focal pathology

Regards
Justin

[thijsdhollander]

Hi Justin,

Thanks for your interest in using SS3T-CSD; that's great to hear!

About the -threshold option, I'm not sure, but I think you might be mistaken what that is used for. I should clarify that the -threshold option actually has no effect at all on the MSMT-CSD (dwi2fod msmt_csd) algorithm, but only applies to the single-tissue CSD (dwi2fod csd) algorithm. It's a value that's used in the mechanism for the non-negativity "constraint" of the original (single-tissue) CSD. It's an iterative optimisation that rather applies a "soft" constraint via regularisation of orientations where (iteratively) values of the FOD are detected that are likely to be zero. This to encourage these values to not dip below zero. You might sometimes hear people talk about "soft (constrained) spherical deconvolution" in this context. The MSMT-CSD algorithm on the other hand uses a "hard" constraint, which doesn't need an artificial threshold to detect values near zero. Finally, SS3T-CSD builds upon the latter hard constraint out of the box too, so the threshold option has no use in this context.
That said, given this is what the -threshold option is used for, even when using the original (single-tissue) CSD, it shouldn't be manipulated for anything else than just getting a robust FOD fit.

So because you mention

to model focal pathology

...I suspect you had a very particular idea in mind? In this context, it's useful to know we've used SS3T-CSD in a range of pathologies as a so-called diffusion "signal representation" (rather than a direct biophysical model of pathologies; which is unlikely to be modelled simply or even extracted from dMRI data due to heterogeneity of pathology).

This was the first work pitching this idea: abstract on ResearchGate

We relied on this in several fixel-based analysis studies to model lesions and pathology inherently, without having to mask them out (including MS, different dementias, stroke, ...). We also looked specifically at using this kind of signal representation to assess heterogeneous tissue composition of lesions:

• First original work doing this for Alzheimer's Disease was this one: preprint on bioRxiv
• We also looked at WMH lesions in stroke using this framework: paper in NeuroImage

Notable in this context is another work looking at WM FODs and tractography in gliomas, hinting at increased sensitivity when using just the highest possible/available b-value: preprint on bioRxiv
(Maybe surprising, but actually quite sensible given this recent work.).

Finally, we've found these 3-tissue compositions to be pretty reliable and stable even long-term: paper in MRM

So all this to essentially say that SS3T-CSD results can be used quite well to fit the diffusion signal even in pathologies, as long as this is approached as a signal representation. I'm wondering whether you wanted to mask out focal pathology, or similar? This question / discussion comes up at times; I was actually speaking to collaborators this morning literally about this topic. 🙂 Let me know if I can help out in other ways, or clarify some of the above for sure!

Cheers,
Thijs

[JustinGarber]

Hi Thijs,

Thanks so much for your thoughtful reply. I obviously misunderstood what -threshold was for.

For my project I am modelling the structural connectome in people with MS. I want to try model the "destructiveness" of MS lesions by having streamlines that track through MS lesions be cut proportional to the lesions "destructiveness". That is for more destructive lesions I want fewer streamlines to track through the lesion, and for less destructive lesions I want more streamlines to track through them. I have been playing around with the -cutoff flag in tckgen, and thought that the -threshold flag would do a similar thing. I have also thought about just relying on SIFT2 to weight the streamlines but have realised that it is not so suited to focal pathology. There might be developments in the future that directly address this http://dx.doi.org/10.1101/2020.01.14.903559

For what it's worth I have a been impressed by SS3T-CSD. My data is single shell b 1000 with 2 x b=0, 64 directions, 2 mm voxels.

Here are some pictures of just anterior to the genu. The background is a 5TT2vis file, with MS lesions in white, grey matter is dark grey, white matter is light grey, CSF is black.

This is MSMT-CSD

This is SS3T-CSD

I think SS3T has done a much better job. But as you can see most MS lesions don't have an appreciable change in ODF amplitude.

Regards
Justin