Which one is better between response msmt_5tt and dhollander for HCP data?

Hi Community,
I obtained two results about response (HCP data) by dwi2response msmt_5tt and dwi2response dhollander, respectively. Which one is better both them?
dwi2response msmt_5tt:


dwi2response dhollander:

Any suggestions there?
Cheers,
Jian

Hi @jiangjiane,

The WM responses from both of those algorithms are typically very similar. The GM responses should be quite similar as well for HCP data (this would be a different story if there were e.g. white matter hyperintense lesions, then dwi2response dhollander is potentially much more accurate). The CSF responses will probably differ somewhat though, even for HCP data. The one obtained from dwi2response dhollander is again more accurate, even for HCP data. I have explicitly tested this myself a while ago: dwi2response msmt_5tt seems to underestimate the b=0 signal and the diffusivity of the CSF response.

Therefore, I would personally advise to use dwi2response dhollander in this scenario!

Cheers,
Thijs

Hi @Thijs,

Thank you for the detailed explanation!
I got it.

Cheers
Jian

1 Like

@jiangjiane, we’ve just enabled a feature that allows you to mark an answer to your question as an “accepted” answer. This will then help future visitor to more easily find questions and existing answers. :slight_smile: See the information here: Community forum: New tags and accepted solutions

Could you try and mark my answer above as accepted (see the instructions in the link I provided)? It would be good to see if the system works. :wink:

@Thijs, I have marked this answer as an “accepted” answer. Thank you again!
I have a question about this problem: Compared with dhollander, there is a bump in the center of the image of msmt_5tt, which results from underestimating the b=0 signal?
Trouble you again, thanks!

Jian

1 Like

Yes, exactly. Essentially, the b=0 and diffusivity of the CSF response get underestimated. Deconvolving such a response leads to overestimation of the CSF compartment (or a “bump” in the CSF, as you say). More importantly, this leads to underestimation of the WM FOD upon partial voluming with CSF or other free fluids. See the original abstract of the “dhollander” method for this specific finding!

Cheers,
Thijs

@Thijs, I have figured out the problem with your help, thank you!

Cheers,
Jian

I have a question about this problem: Compared with dhollander, there is a bump in the center of the image of msmt_5tt, which results from underestimating the b=0 signal?

I think the ‘bump’ you are referring to here is that in the top-right image, i.e. the WM response function for the maximal b-value? In which case: This isn’t to do with diffusivities, or indeed the b=0 image at all, but manifests when the DWI signal from multiple single-fibre WM voxels are rotated and combined together to produce a single WM response function. We used to see this sort of effect with much more prevalence when using the old FA > 0.7 single-fibre selection heuristic. It’s probably a combination of using voxels that are not ‘filled’ with WM fibres, voxels that do contain some small crossing-fibre bundles, that sort of thing.

Both dwi2response msmt_5tt and dwi2response dhollander in fact use dwi2response tournier internally to go from a WM mask to a WM single-fibre mask. So there must be something about the conservative-ness of dwi2response dhollander’s initial WM mask that helps to prevent using voxels in the WM response function estimation that dwi2response msmt_5tt used but perhaps shouldn’t have.

Additional point: As of Release Candidate 1, dwi2response algorithms use a new mechanism for performing the combination of single-fibre DWI signals, which explicitly forbids the formation of precisely this type of ‘bump’.

Rob

Hi Rob,
Thank you for such a detailed answer.

Cheers,
Jian