5tt multi-shell response functions - Global tractography

Dear experts,
I am a beginner on MRtrix, interested in global tracktography. I am trying to figure out how
to create multi-shell response functions without T1 data. Is there an alternative way to
create 5tt.mif ?
Please advice, Thanks, Carlos.

Hi Carlos,

Nice to hear you’re interested in global tractography. The issue of response function (RF) estimation without T1 is certainly a very relevant one, and one that’s also discussed in section 5.2 of the paper.

The core issue is to select reliable voxel masks for CSF, GM, and single-fibre WM, in which the RFs are subsequently estimated. The main advantage of having a T1w-image is that it provides a high degree of confidence that the segmented voxel masks indeed correspond with WM, GM, and CSF tissues, thus supporting the reliability of the estimated RFs.

If no T1 is available, you could think of some other heuristic to segment “pure” WM, GM, and CSF voxels in the DWI data directly. An example of such approach is the dhollander method in dwi2response, which is very user friendly because you don’t need the 5tt segmentation. However, the selected voxel masks may not always provide the same level of trust as a T1 segmentation, especially in pathology or preclinical data.

Another approach that I personally find more interesting, is to estimate RFs and ODFs simultaneously from the DWI data in a way that maximally supports the underpinning model assumptions in MT-CSD. In recent work we developed a method to do exactly that: a fully unsupervised decomposition of DWI data into convolutional components associated with WM, GM, and CSD in the healthy adult brain, but also e.g. with edema in pathology. Unfortunately this is not available in MRtrix yet, but if you get in touch I can set you up.

I hope that answers your question. If not, let us know.


In quite a few pathological cases (including extensive white matter hyperintense lesions), the “dhollander” algorithm performs actually much more trustworthy than the 5ttgen fsl output driven dwi2response msmt_5tt approach; due to 5ttgen fsl misclassifying some of those common pathologies as a certain tissue type. While initially not even intended, we’ve found that the “dhollander” algorithm happens to inherently cope with those common pathologies surprisingly well. :smile: For animal data of certain specific kinds, tuning might be needed though.

@Carlos_Gutierrez, this should do the trick:

dwi2response dhollander dwi.mif response_wm.txt response_gm.txt response_csf.txt

…where dwi.mif is your (pre-processed) DWI dataset. No T1-weighted image needed!

If you’d like to know more about how this method works, give this abstract a good read.

@ThijsDhollander, thanks for sharing the abstract. Is the algorithm implemented in MRtrix?

We refer to this method as “single-shell 3-tissue CSD” (SS3T-CSD).

I’m assuming it would be in dwi2fod, but it’s not listed (but I may have missed it).

Hi @blezek,

The SS3T-CSD would be in dwi2fod if it were available… but it’s not available yet indeed. The above linked abstract itself however, is on the unsupervised response function selection algorithm that is available in dwi2response, more specifically as the “dhollander” algorithm choice in that command/script. This algorithm’s goal is specifically to obtain the WM, GM and CSF responses directly from the DWI data. It works for both multi-shell as well as single-shell (+b=0) data. In case of “true” multi-shell data (3 distinct b-values, including b=0), the responses are then suitable for MSMT-CSD (dwi2fod msmt_csd). For single-shell (+b=0) data, you’d have to input them into SS3T-CSD (once I make it available at some point in the hopefully-not-too-distant future). Alternatively for such data, you can ditch the GM response and do an MSMT-CSD with only 2 tissue types (WM and CSF) (happily read the rest of that thread to get as much context as needed). You don’t get to enjoy the benefits of a GM response then of course. Here’s some further information on dwi2response and dwi2fod algorithms, and some relations and potential combinations.

In the context of the original question of this thread though: @Carlos_Gutierrez seems to have multi-shell data, which would work well the global tractography algorithm available in tckglobal. This is a multi-shell multi-tissue global tractography algorithm, that works directly on the DWI data (no explicit CSD step of any sorts, but the same “model” and assumptions with respect to the DWI signal are inherently used and relied upon in the method). With multi-shell data (3 distinct b-values, including b=0), that method could generate a track(segment) distribution that represents the WM, in addition to GM and CSF isotropic signal compartments. In the absence of multi-shell data, it can (similar to MSMT-CSD theory here) also just go for the WM track distribution together with only a CSF compartment. Again, you wouldn’t benefit of all that a GM response can bring additionally in ensuring a good as well as meaningful fit.

But so, in @Carlos_Gutierrez’s case, he can benefit of the dwi2response dhollander algorithm to get his multi-shell WM, GM and CSF responses, which can then feed into tckglobal.

I hope this all makes sense. Don’t hesitate to ask for clarifications if it doesn’t. :wink:


Dear Daan (Carlos, and Thijs),

I am a clinician uses tractography routinely for presurgical planning. Much of my work involves performing tractography using brain tumor/epilepsy lesion data. I found this question
of using (T1-less) pure DWI data-driven, tissue RF estimates, extremely relevant to my line of work. My experience is that more often than not T1-based tissue classification is erroneous over the pathological region, thus render MSMT-CSD unreliable.

In our centre, we work with both single shell b3000 data and more recently multiband, multishell DWI data (b = 1000, 2000, 2800).

DWI data-driven approach offers an alternative. After reading this post, I am also interested in your recent paper (Christiaens, NI, 2017), and would like to try adopt it to my work flow.

My crude understanding of this technique (I am a clinician): Decomposes DWI into tissue specific ODFs and corresponding RFs (rather than having a priori tissue RFs as basis of most fODF reconstructions within MRtrix). This is done using the nonnegativity constrained NMF. Because no a priori RF is required, it can be done in a fully unsupervised manner.

I have a question regarding number of components estimated from the method,if I suspect more than one pathology existed in the region of interest: A case in example is illustrated in Figure 10 of your paper- post-high grade glioma resection, with peri-lesional edema. There are clear evidence of surgical cavity rim enhancement, representing either tumour residual or post-resection gliosis. So, If I have set the number of components to five empirically, can I expect one pathological ODF correspond to perilesional edema, and another correspond to tumour residual?

I should also mentioned that I have been experimenting with Thijs’ SS3T-CSD method using the HARDI-shell data (Well, to be more exact: “WM+CSF” two-tissue RF+ msmt-csd). It does “clean-up” CSF FODs quite a bit, but will wait for the fully release version of it to evaluate the technique more properly in these clinical settings.

Many thanks,


Hi Joseph,

You should probably take a look at this abstract I’ll be presenting in a couple of weeks at ISMRM: https://www.researchgate.net/publication/315836029_Towards_interpretation_of_3-tissue_constrained_spherical_deconvolution_results_in_pathology . Note how the FODs are cleaned up in the lesions. Key to this, however, is the “GM” compartment.

This would hold reasonably well for both MSMT-CSD as well as SS3T-CSD (when SS3T-CSD becomes available of course), for the cases where each of those uses WM, GM and CSF compartments. Where we’re going towards with that is to refer more to them as “WM-like, GM-like and CSF-like”.

This is separate from the issue of what data is available. For single-shell data, you’d need SS3T-CSD (as is done in that abstract; but again, this requires the SS3T-CSD implementation of course), but having multi-shell data in your case, MSMT-CSD could be a good option. This then does depend on at least 3 distinct b-values having good data. You’ve got 4 b-values (0, 1000, 2000, 3000) for the multi-band data, so you can afford to drop one if needed (I understand that motion, in your case, is likely to sometimes mess up an entire shell’s data, correct?).


Hi Joseph,

Nice to hear that you’re interested in using our work on constrained spherical factorisation.

That’s correct. CNSF decomposes DWI data into a number of components, each of which is a spherical convolution of a response function and an ODF. RFs and ODFs are estimated simultaneously, under nonnegativity constraints of the ODFs and convexity constraints of the RFs (i.e., they must occur somewhere in the data).

What makes this an unsupervised technique, is that the number of components is a free parameter that can be chosen by the user (instead of being fixed to WM-, GM- and CSF-like compartments). This is conceptually similar to blind source separation, such as ICA in fMRI. While CNSF is thus not constrained to decompositions in WM, GM, and CSF, it’s encouraging to see healthy brain is nevertheless factorised into components associated with these 3 tissues: it shows that Ben and Donald made a clever choice for their 3-tissue model in MT-CSD :slight_smile: Preclinical in vivo rodent data is also directly decomposed into WM/GM/CSF-associated components.

Because CSNF can work with any number of components (and SH orders of each of them), selecting that number is indeed the main question when using the method. Until I have a quantitative criterion, I advise users to hypothesise a few different candidates and then compare their results. However, a main limitation is that I advise the number of components to be less than or equal to the number of b-values in the data, because these provide the main contrast to separate tissue components. Hence, for your multi-shell data, the advised number of CNSF sources is limited to 4.

I agree that it would be amazing if perilesional edema, tumour residual, and potentially even tumour substructure, could be discriminated in different components. Unfortunately, I don’t have glioma patient DWI data with more than 4 shells. In the 4-shell data that I looked at, the 4th component was generally associated with edema, with variable robustness depending on data quality. I would love to have very high-quality high-multi-shell patient data to explore this further.

I hope that answers your question. Get in touch if you like to discuss this further.

Best regards,

Thanks for the abstract, Thijs,

The technique definitely looks promising.
We can discuss more at ISMRM next week.

Looking forward to your talk


Dear Daan,

Thanks for the detailed response.
I am keen to explore your technique.
Will you be at ISMRM next week? perhaps we should catch-up to discuss more in detail

Thanks again,


Hi Joseph,
Yes, I will be at ISMRM. Looking forward to seeing you there.