CSF voxel selection in dwi2response dhollander skips ventricles, expected?

[AndreasP]

Greetings MRTrix community and experts! This will be my first post, first of many I would assume .

I want to start of by giving some background about my knowledge of the area: at the moment of writing this I have finished my 4th year as a master student in engineering physics although I have taken some courses in biomedical engineering and read some papers regarding this topic, thus my knowledge of the brain is fairly limited.

Nevertheless, we all start somewhere and this far I have completed the MRtrix tutorial provided on Andy’s Brain Book and I am now in the middle of trying to replicate the process on data for a research project regarding MRgFUS that I am working on, thus the DRTTs are the tracts we at the moment are trying to delineate.

The DWI-data is gathered on a 3T GE-scanner, it is multi-shelled with shells: b=0, b=1500, b=3000 [s/mm²] of dimensions 256x256x82x102 with voxel size 1x1x1.8x4.961, I am using MRtrix version 3.0.5.

Now regarding my question, I have estimated the RFs using the “dwi2response” command with the “dhollander” algorithm, running the command:
$ dwi2response dhollander DTI99_den_preproc_unbiased.mif wm.txt gm.txt csf.txt -voxels voxels.mif

I include below the information provided by the command regarding the process

dwi2response: Computing brain mask (dwi2mask)…
dwi2response: -------
dwi2response: 3 unique b-value(s) detected: 0,1500,3000 with 9,47,46 volumes
dwi2response: -------
dwi2response: Preparation:
dwi2response: * Eroding brain mask by 3 pass(es)…
dwi2response: [ mask: 997220 → 819087 ]
dwi2response: * Computing signal decay metric (SDM):
dwi2response: * b=0…
dwi2response: * b=1500…
dwi2response: * b=3000…
dwi2response: * Removing erroneous voxels from mask and correcting SDM…
dwi2response: [ mask: 819087 → 817679 ]
dwi2response: -------
dwi2response: Crude segmentation:
dwi2response: * Crude WM versus GM-CSF separation (at FA=0.2)…
dwi2response: [ 817679 → 480619 (WM) & 337060 (GM-CSF) ]
dwi2response: * Crude GM versus CSF separation…
dwi2response: [ 337060 → 247876 (GM) & 89184 (CSF) ]
dwi2response: -------
dwi2response: Refined segmentation:
dwi2response: * Refining WM…
dwi2response: [ WM: 480619 → 437185 ]
dwi2response: * Refining GM…
dwi2response: [ GM: 247876 → 166875 ]
dwi2response: * Refining CSF…
dwi2response: [ CSF: 89184 → 56748 ]
dwi2response: -------
dwi2response: Final voxel selection and response function estimation:
dwi2response: * CSF:
dwi2response: * Selecting final voxels (10.0% of refined CSF)…
dwi2response: [ CSF: 56748 → 5675 ]
dwi2response: * Estimating response function…
dwi2response: * GM:
dwi2response: * Selecting final voxels (2.0% of refined GM)…
dwi2response: [ GM: 166875 → 3338 ]
dwi2response: * Estimating response function…
dwi2response: * Single-fibre WM:
dwi2response: * Selecting final voxels (0.5% of refined WM)…
dwi2response: Selecting WM single-fibre voxels using built-in (Dhollander et al., 2019) algorithm
dwi2response: [ WM: 437185 → 2186 (single-fibre) ]
dwi2response: * Estimating response function…
dwi2response: -------

The RFs looks reasonable to me, the volumes are shaped as I would expect with the WM showing a clear anisotropic behavior for non-b0 values and GM/CSF being isotropic, their magnitudes decreasing with increasing shells. I would provide an image but it seems as a new user I could only include one and I believe the RFs can be described by their parameter values instead:

WM:
4801.50950503143 0 0 0 0 0
2275.33020633093 -1125.49358817565 371.228823973852 -79.9636359452689 8.53893164983394 0.722755748079911
1604.07770671768 -932.78442680307 503.157969439236 -195.247110095876 52.3921839876302 -8.5322801271091

GM:
7769.72345594421
2508.93076132426
1080.33320268872

CSF:
16697.4317257711
303.518197070762
241.140721262514

Now my question is regarding the chosen voxels for the estimation, I provide the image below overlayed on the b0 DWI image:

and that which seems strange to me is the CSF voxels, specifically that none seems to be taken from the ventricles, which to my understanding should be the location of most representative CSF voxels of free diffusion. As I have understood, the dhollander algorithm leverages no anatomical information but instead chooses voxels based on how well they represent the expected characteristic diffusion signal of that tissue. So maybe this is not actually a problem since the RFs looks good, but I wonder what it says about my data if the most representative CSF voxels are not located in the ventricles, is it down to partial-volume-effects, or is it expected behavior?

Kind regards,
Andreas

[AndreasP]

I think I found a good explanation here regarding the voxel placement of the dhollander algorithm!

Still my tractography results are poor (the streamlines cover little of the brain and are not especially coherent), I believe it to be related to the FODs, I found this discussion here about how the combination of MSMT and ACT (which I am using) likely will require a lowering of the FOD amp. threshold, I went as low as 0.045 with tractography results still poor.

I generated a seed density map to inspect the seeding process:

$ tckmap tracks_10M.tck -template DTI99_den_preproc_unbiased.mif seed_density.mif

$ mrview DTI99_den_preproc_unbiased.mif -overlay.load seed_density.mif -overlay.colourmap 2

and it has a maximum intensity of only ~5700, if I run the same command on the tractography I generated from Andy’s Brain Book tutorial, the intensity is much stronger in the entire brain with a max of ~20000 instead. Thus my seed mask being vastly underpopulated could be one of the reasons for the poor result, maybe gm/wm boundary-seeding mask generated for the ACT is poor and is limiting the seeding which in turn limits the tractography.

I then checked the FOD amplitudes as the following:

$ mrmath wmfod_norm.mif rms -axis 3 fod_amp.mif

for both my data and the tutorial, and both had a maximum intensity of 0.160, which to me suggest the error does not lie here, I include the FOD amplitude map from my data just as to visualize what I am talking about.

Regarding the FODs them-self I include a zoomed in image here (in the region of what I think is the “splenium of corpus callosum” after a quick search ):

and I might not know much of the tracts present in the brain, but to me it seems that the FODs here, specifically (but not limited to) the FODs representing posterior-anterior direction are orientated completely orthogonal to the direction I would expect, it just makes no sense to me and likely is the main culprit of my poor tractography (I want to make clear that I also checked this before attempting the tractography, as that is always good to do, and it caught my eye already back then but since it was my first time I wanted to perform the tractography still and see the results).

I.e. the FODs is the first thing I now will attempt to fix, however I am not quite sure how to actually do this, I have some ideas as what might be the problem and I will try them, I am however still very much a novice to MRtrix and thus appreciate any guidance/suggestion.
I did not expect it to be smooth ride considering this is my first experience with this area, and I suppose this is how you learn, the positive is that there is a lot of room for improvement!

Kind Regard,
Andreas

[bjeurissen]

Clearly your ODFs are pointing in the wrong way! This probably means that DTI99_den_preproc_unbiased.mif does not contain the correct gradient directions. You can check (and even correct this) using dwigradcheck DTI99_den_preproc_unbiased.mif. This should inform you that the gradient directions stored in your file are not correct. Rather than using the corrected gradient directions provided by the command, it is probably best to figure out where in your pipeline this error was introduced and fixing it.

See Automated correction of improperly rotated diffusion gradient orientations in diffusion weighted MRI - PubMed for a detailed explanation of the problem and rationale of dwigradcheck.