[Mrtrix-discussion] Response functions at lmax=6 vs lmax=8

Mon Apr 4 12:24:45 PDT 2016

Thanks Donald for those references. There are some very nice insights in there.

-David

From: J-Donald Tournier [mailto:jdtournier at gmail.com]
Sent: Sunday, April 03, 2016 4:24 PM
To: David Grayson <dgrayson at ucdavis.edu>
Cc: mrtrix-discussion at www.nitrc.org <mrtrix-discussion at public.nitrc.org>
Subject: Re: [Mrtrix-discussion] Response functions at lmax=6 vs lmax=8

Hi David,

So I guess this is where the issue of quantification becomes a matter of debate. As was shown quite comprehensively by Derek Jones et al<http://www.ncbi.nlm.nih.gov/pubmed/22846632>, streamlines counts are not in themselves particularly meaningful - and hence can't be viewed as quantitative. There is also evidence that they're not particularly reproducible (at least when using probabilistic streamlines without an enormous number of streamlines) - see e.g. our TDI paper on the issue<http://www.ncbi.nlm.nih.gov/pubmed/24246491>. This is what I meant when I said there isn't much quantification you can do in MRtrix 0.2. You have a lot more options in MRtrix3, including ACT<http://www.ncbi.nlm.nih.gov/pubmed/22705374>, SIFT<http://www.ncbi.nlm.nih.gov/pubmed/23238430>, SIFT2<http://www.ncbi.nlm.nih.gov/pubmed/26163802>, and AFD<http://www.ncbi.nlm.nih.gov/pubmed/22036682> - SIFT being particularly beneficial for connectomic analyses<http://www.ncbi.nlm.nih.gov/pubmed/25312774>.

As to your question: I don't think these minor inaccuracies in the responses are going to make much difference at all to the tractography itself. These are minor noise ripples in the low signal region of the response - what makes a big difference is the overall size of the response (this influences the scaling of the FOD), and how broad it is. But even if it does have a minor effect on the FOD, this is unlikely to affect the tractography a great deal, and is even less likely to influence any comparisons you might make between subjects - I don't see how this could introduce a bias between your groups, at worst you might lose a negligible amount of statistical power. This of course all assumes that the procedure you're using is the same for all subjects.

So I would run at lmax=8 without worrying too much about it... If you were to switch to use some of the methods in MRtrix3, then this would indeed be a different matter.

Cheers,
Donald.

On 1 April 2016 at 18:02, David Grayson <dgrayson at ucdavis.edu<mailto:dgrayson at ucdavis.edu>> wrote:
Hey Donald,

I am in fact doing quantitative analyses of number of pobabilistic streamlines connecting a given pair of regions, across subjects. So I would prefer to minimize variation in the rf across subjects.

From your experience, would you expect one of these rf's to be more consistent across subjects?
________________________________
From: J-Donald Tournier <jdtournier at gmail.com<mailto:jdtournier at gmail.com>>
Sent: Friday, April 1, 2016 12:15:30 AM
To: David Grayson
Cc: mrtrix-discussion at www.nitrc.org<mailto:mrtrix-discussion at www.nitrc.org>

Subject: Re: [Mrtrix-discussion] Response functions at lmax=6 vs lmax=8

Hi David,

Yes, these look fine. That's often a little protuberance like this in the response, it'll be due to some extent to the Rician bias driving the signal away from zero in this region. It's a minor inaccuracy that shouldn't make a great deal of difference to the output of CSD. In fact, some people reckon it provides a better fit to the actual signal (since it contains Rician bias anyway), and yields sharper FODs... Not entirely sold on that one, but just to say that it really shouldn't impact on any downstream inferences.

On that note, these subtle differences and inaccuracies in the response may conceivably make a difference if you're doing something quantitative with the tracks themselves across subjects. There's really very little you can do on that front with MRtrix 0.2. If you'd been using MRtrix3, there would be many more options available for quantification, at which point you'd need to worry about this. But even then, the main concern is to make sure the same response is used across subjects, rather than whether the response is perfect...

Hope this helps,
Donald

--
Dr J-Donald Tournier (PhD)

Senior Lecturer, Biomedical Engineering
Division of Imaging Sciences & Biomedical Engineering
King's College London

A: Department of Perinatal Imaging & Health, 1st Floor South Wing, St Thomas' Hospital, London. SE1 7EH
T: +44 (0)20 7188 7118 ext 53613
W: http://www.kcl.ac.uk/medicine/research/divisions/imaging/departments/biomedengineering
On 1 Apr 2016 4:39 am, "David Grayson" <dgrayson at ucdavis.edu<mailto:dgrayson at ucdavis.edu>> wrote:

Thanks for the responses here.

I should have mentioned, I have been using the freesurfer WM mask (co-registered to the B0) to exclude noisy voxels. Looks like there was an issue with co-registration. Now that this is fixed, indeed the lmax=8 response looks (very slightly) sharper than lmax=6.

However, I'm wondering if you can have a look at the following comparisons of lmax=6 vs lmax=8, using an eroded WM mask and using the full WM mask. The lmax=8 still produces the protruberance on the z-axis, which is exaggerated using the eroded WM mask. Does this suggest that I should perhaps stick with lmax=6, and avoid eroding any further the WM mask? Would you consider these RF's acceptable now?

http://imgur.com/42NCktt

Thank you very much,

David

________________________________
From: albnet at gmail.com<mailto:albnet at gmail.com> <albnet at gmail.com<mailto:albnet at gmail.com>> on behalf of Dorian P. <alb.net at gmail.com<mailto:alb.net at gmail.com>>
Sent: Tuesday, March 22, 2016 3:10 PM
To: J-Donald Tournier
Cc: David Grayson; mrtrix-discussion at www.nitrc.org<mailto:mrtrix-discussion at www.nitrc.org>
Subject: Re: [Mrtrix-discussion] Response functions at lmax=6 vs lmax=8

I always used to correct manually the single fiber mask. The automatic thresholding includes out of brain voxels in inferior slices. A simple threshold is not to be trusted, I think.

Dorian

On Tue, Mar 22, 2016 at 6:01 PM, J-Donald Tournier <jdtournier at gmail.com<mailto:jdtournier at gmail.com>> wrote:
Hi David,

I've never heard of this being referred to as a 'nubbin', but it's clear enough... ;)

Looking at your responses though, the 'nubbin' isn't all that worrying. What's more worrying is the fact that the lmax=6 response is sharper than the lmax=8 response, which seems wrong. Your lmax=8 response is certainly broader than I'd expect at b=3000.

So all this suggests that the response function estimation isn't working all that well, which is typically a symptom of a poor single-fibre mask. You used an FA threshold of 0.7 here, but bear in mind that this is by no means a 'default' - this is more of a guideline, as stated in the documentation<http://jdtournier.github.io/mrtrix-0.2/tractography/preprocess.html#csd>. Here's the relevant excerpt:

Note that this value is a guide only - feel free to use a different value if this does not produce satisfactory results. Ideally, you should now have a mask containing a few hundred voxels, all located within high FA white matter regions. It is very important to check that the single-fibre mask is suitable, as otherwise the response function produced in the following step may be totally inappropriate, which would seriously affect the quality of the CSD output. If needed, you can edit this mask image to remove unwanted voxels using the ROI analysis<http://jdtournier.github.io/mrtrix-0.2/general/mrview.html#roi> sidebar tool within MRview<http://jdtournier.github.io/mrtrix-0.2/general/mrview.html>.

I would have a good look at the single-fibre mask used in this step, and check against the above. The chances are you might find a lot of noisy high-FA edge voxels got included or something, despite the erosion step. This all depends on how good the initial brain mask was (I often found it difficult to exclude the nasal sinuses, for example). Unfortunately, this step can be a bit fiddly.

Alternatively, you could upgrade to MRtrix3<http://www.mrtrix.org> - we've put in quite a bit of work on estimating the response function<http://mrtrix.readthedocs.org/en/latest/concepts/response_function_estimation.html>...

All the best,
Donald.

On 22 March 2016 at 19:47, David Grayson <dgrayson at ucdavis.edu<mailto:dgrayson at ucdavis.edu>> wrote:
Hi MRtrixers,

I have HARDI datasets on a large sample of young adults (7-16yrs) taken with 72-dir and b0=3000. I am wondering whether I should proceed with lmax=6 or 8. I am using MRtrix 0.2.12.

I am concerned about using lmax=8 because in the response function there is often a ‘nubbin’ that appears right in the center of the Z-axis line that doesn’t appear at lmax=6. The below link shows an example of the RF’s generated for a single subject taken at lmax=6 (on the left) and lmax=8 (on the right). These are generated using default parameters (FA threshold of 0.7 for the single-fiber-orientation mask).
http://imgur.com/a/ODrcf

Hopefully it’s obvious what I’m talking about. Sometimes this nubbin is more pronounced than what I’m showing here. I’m not sure how concerned I should be about this, but based on my intuition about what the RF means, the Z-axis itself should represent the global minimum of the DW signal, which is obviously not the case with lmax=8. Should I avoid any appearance of this effect at all costs, or is there some rule of thumb for how big it can be before I should be concerned about it?

Thanks,

David

_______________________________________________
Mrtrix-discussion mailing list
Mrtrix-discussion at www.nitrc.org<mailto:Mrtrix-discussion at www.nitrc.org>
http://www.nitrc.org/mailman/listinfo/mrtrix-discussion

--
Dr J-Donald Tournier (PhD)

Senior Lecturer, Biomedical Engineering
Division of Imaging Sciences & Biomedical Engineering
King's College London

A: Department of Perinatal Imaging & Health, 1st Floor South Wing, St Thomas' Hospital, London. SE1 7EH
T: +44 (0)20 7188 7118 ext 53613<tel:%2B44%20%280%2920%207188%207118%20ext%2053613>
W: http://www.kcl.ac.uk/medicine/research/divisions/imaging/departments/biomedengineering

_______________________________________________
Mrtrix-discussion mailing list
Mrtrix-discussion at www.nitrc.org<mailto:Mrtrix-discussion at www.nitrc.org>
http://www.nitrc.org/mailman/listinfo/mrtrix-discussion

_______________________________________________
Mrtrix-discussion mailing list
Mrtrix-discussion at www.nitrc.org<mailto:Mrtrix-discussion at www.nitrc.org>
http://www.nitrc.org/mailman/listinfo/mrtrix-discussion

--
Dr J-Donald Tournier (PhD)

Senior Lecturer, Biomedical Engineering
Division of Imaging Sciences & Biomedical Engineering
King's College London

A: Department of Perinatal Imaging & Health, 1st Floor South Wing, St Thomas' Hospital, London. SE1 7EH
T: +44 (0)20 7188 7118 ext 53613
W: http://www.kcl.ac.uk/medicine/research/divisions/imaging/departments/biomedengineering
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.nitrc.org/pipermail/mrtrix-discussion/attachments/20160404/b417e103/attachment-0001.html>