SS3T-CSD on b=1000, 31 gradients directions, one b0, non isotropic resolution

[atefbadji]

Hello, I have a single shell (one b0 and 31 b1 volumes) data with low b-value (b1=1000), 31 directions. Non-isotropic resolution: 2.0 x 2.0 x 2. 4 mm . I applied common pre-processing steps (https://3tissue.github.io/doc/single-subject.html): denoising, unringing, motion and eddy current distortion correction and bias field correction. I didn't do an upsampling or regridding of the data for the purpose of this quick test. I estimated the response function of each tissue (WM, GM, CSF) using the dwi2response dhollander command.

3 Tissue response function estimation (reports of SDM)

Color code: WM = blue; GM=green; CSF=red
The signal decay metrics (SDM) increase from WM to GM to CSF

SS3T-CSD tissue compartment images (GM, WM, CSF)

Then I performed a 3 CSD tissue modeling for single-shell data (SS3T-CSD)

Although the result may not look "clean", it seems that the quality is a bit better at the posterior end of the brain compared to the anterior regions as shown in previous reports. Ideally, I believe that one would want that the GM-like signal is filtered out as much as possible to reduce the bias in the estimation of the WM-like signal. How can we assess that the GM-like signal is properly filtered out?

SS3T FOD images (FOD-DEC map)

Then, I computed an FOD-based DEC map from the WM FOD image

Tractography (cut-off 0.07)
Finally, I performed probabilistic tractography using the SS3T setting (cutoff 0.07)

Tractography (cut-off 0.09)

How to determine the appropriate cut-off to stop tracking appropriately ? It seems to be a qualitative assessment, a trade-off between false positive and false negatives.

Thank you for any feedback on my results and for your very well-written and thorough documentation on your website.
Best,
A

[thijsdhollander]

Hello, I have a single shell (one b0 and 31 b1 volumes) data with low b-value (b1=1000), 96 directions. Non-isotropic resolution: 2.0 x 2.0 x 2. 4 mm.

Nice to see it being tested on another very "real world" clinical scenario. The added anisotropic resolution is neat (as a challenge for the pipeline) here; but I'm pretty confident the dwi2response dhollander can handle it. 😉

A short question about the data maybe: you mention "31 b1 volumes" but also "96 directions". I'm slightly confused, as the number of b1 volumes is typically what's "also" called the number of gradient directions. Do you mean the protocol has a few repetitions of this, for example 3 repetitions of a 1 b=0 and 31 b=1000 protocol?

I didn't do an upsampling or regridding of the data for the purpose of this quick test.

Yep, that's entirely sensible. 👍 For a lot of purposes, you might not really need upsampling at all in general. For data with an anisotropic spatial resolution, I would eventually perform some kind of upsampling, or just "resampling" in general though. Note that, just like the pipeline shows, this only has to be done right before the SS3T-CSD step. All other steps before that, including dwi2response dhollander, work perfectly fine (or even better!) without upsampling/resampling the data; those steps are all designed to run directly on the original resolution of the data. If you'll eventually be processing more datasets with this protocol for e.g. group analysis or other purposes, I'd recommend resampling to e.g. "at least" 2x2x2 mm^3 isotropic voxels. You can also go further to e.g. 1.5 mm isotropic voxels, but that will of course start to impact the computation time for SS3T-CSD more, as there are then way more voxels to compute a 3-tissue CSD model for.

I estimated the response function of each tissue (WM, GM, CSF) using the dwi2response dhollander command.

Looking all very good! The voxel selections on the image you present look 100% sensible, just as they should be. A visualisation tip: you can switch off the "interpolation" of the overlay (in the overlay tool) if you're overlaying those red / green / blue voxels: that will more clearly show the exact "discrete" voxels, without those black kind of boundaries (due to interpolation). Some of those you now see, might be coming from the previous or next slice in the 3D image. It also looks a bit "cleaner" (and more "professional" 🤓 ) to present in slides or publications.

As I mentioned before on Twitter, you can potentially also plot the b=0 to b=1000 "decay" for WM, GM and CSF responses: for GM and CSF, it's just those 2 numbers, for WM it's the first number of each row. The other numbers for WM encode the "disk" shape of the response function: that's useful for shview, which can show the full profile in 3D. Other than that, those other numbers for WM aren't "obviously" intuitive for most people. If you were to plot the WM, GM and CSF "decay" functions (just a point for b=0 and b=1000 connected by a straight line here), you would indeed also "see" their relative signal decay (metric) or SDM behaviour: the CSF line would be pretty steep, the GM line would be a lot flatter, and the WM line would be still (slightly) more flat than GM. For the SS3T-CSD algorithm in particular in the next step, the command will also check and even report the SDM values at the command line.

Then I performed a 3 CSD tissue modeling for single-shell data (SS3T-CSD).

Again looking very neat! The quality is very much in line with what's now expected, given some of the feedback posted by others before. This is really nice to see: the pipeline proves to be very reliable indeed. All 3 compartments come out very nicely. Note some interesting patterns posterior to the ventricles, with elevated GM-likeness and CSF-likeness (free-water) showing distinct (unique) patterns within the white matter.

Although the result may not look "clean", it seems that the quality is a bit better at the posterior end of the brain compared to the anterior regions as shown in previous reports. Ideally, I believe that one would want that the GM-like signal is filtered out as much as possible to reduce the bias in the estimation of the WM-like signal. How can we assess that the GM-like signal is properly filtered out?

That's a difficult but relevant question indeed. On a basic level, it just has to look (qualitatively) sensible of course: you want to see GM-like signal in the cortical gray matter for example. Similarly, you want the WM-like signal to be clearly lowered in the cortical gray matter, relative to the neighbouring white matter. Without a 3-tissue model and at such a low b-value for example, that would already (almost or entirely) not be the case anymore. In your result, we can see the WM-GM boundary nicely depicted in the WM-like compartment image, with the white matter region looking much brighter than the gray matter region. That's basically the sign you're looking for.
Other than that, the story gets more complicated, as the exact result will depend on b-value; that's also completely normal and expected! SS3T-CSD for a given b-value (single-shell + b=0) is / should not be expected to be the same as for another b-value or as for MSMT-CSD with multiple b-values. None of these is in general more correct (no gold standard exists here); that's why I introduced the "WM-like", "GM-like" and "CSF-like" terminology, and its exact biological or physical definition is b-value dependent. Coincidentally, Sila Genc et al. just uploaded a relevant preprint in this context: https://www.biorxiv.org/content/10.1101/2020.01.15.905802v1.full . The question of "more accurate" only applies if you're after a very specific interpretation. If that interpretation would be "apparent fibre density", which aims to be more specific to the intra-axonal signal, then Genc et al. show that a higher b-value will get you closer to that; lower b-value(s) are more contaminated with extra-axonal signal and thus "contaminate" the WM-like signal indeed more with extra-axonal signal. This will likely reduce the contrast between tissue types' signal also.

...but putting all of that aside (as it can't be truly "validated" or quantitatively checked for your result), it's mostly about checking that it looks sensible. With regard to "want that the GM-like signal is filtered out as much as possible"; a sort of general warning: yes, that is true, but that doesn't per se mean that the GM-like signal should be as high as possible of course. You want the "true" GM-like signal to be filtered out as much as possible. You could even argue that, to be on the "careful" side, you'd rather want a bit less GM-like signal filtered out instead of too much (false) GM-like signal filtered out: if the GM-like signal is overestimated, you might otherwise lose evidence of real (anisotropic!) WM-like signal. Note this work for example: https://www.biorxiv.org/content/10.1101/629873v1.full . We found that MSMT-CSD can have a tendency to over-estimate GM-like signal in particular scenarios, e.g. in the gliomas shown in that work; and this then comes at the cost of losing valuable WM FODs that indicate a remaining presence of (infiltrated) axon bundles. We're still working on an improved version of that preprint, based on feedback, to include more explanation providing a better insight / intuition as to why this can effectively be the case.

To conclude this bit in a reassuring way though: given the other feedback as well as your result, it looks like the default parameters for the SS3T-CSD do provide you with such a "safe" or "conservative" estimate of the 3-tissues. So generally, there should (hopefully) be little to worry about here. 👍

Then, I computed an FOD-based DEC map from the WM FOD image.

Nice! Yes, this is the best way to visualise the WM compartment as an image/map; it uses its intensity as well as the colours as much as possible to visualise all relevant aspects. The FODs (as an overlay, as you show) provide further details. Note again the nice WM-GM boundary depiction also. The FODs themselves are of quite amazing quality, given the properties of your dataset. Great news!

Finally, I performed probabilistic tractography using the SS3T setting (cutoff 0.07).

How to determine the appropriate cut-off to stop tracking appropriately ? It seems to be a qualitative assessment, a trade-off between false positive and false negatives.

Yep, 100% correct. For lower b-values, the balance will likely also be harder to find, although 3-tissue CSD modelling already helps a great deal in making this easier and better. As far as I can "guess" based on the results you show, I reckon the 0.09 cutoff is better here. It clearly removes some false positives. The risk in trying to assess this though is that a lot of the "first" false negatives that would pop up when increasing the cutoff are very hard to spot in a dense whole-brain result. The cutoff probably differs between parts of the brain and tracts/bundles even; so no single cutoff is perfect... it really is a trade-off for whole-brain tractography! If you're performing targeted tractography to segment a specific bundle though, it might become more straightforward to assess the result and determine an appropriate tweaking of the cutoff. Note how in the aforementioned work ( https://www.biorxiv.org/content/10.1101/629873v1.full ) we even introduced a mechanism to tweak the cutoff locally in the tumour region, as the presence of healthy WM was substantially lower (but still present and thus highly relevant for surgery!) in that region.

Thank you for any feedback on my results and for your very well-written and thorough documentation on your website.

No worries at all, and thank-you! I'm now convinced little to no tweaking is actually needed, even for those low b-values. When I find some time, I'll update the information on https://3tissue.github.io/doc/ss3t-csd.html to reflect this. When it was initially written, I hadn't tested the current (internal) mechanism that SS3T-CSD uses to balance b=0 and b=... data yet that extensively; but judging from the feedback, it looks like the default parameters of that mechanism result in a robust behaviour across a wide range of b-values out of the box. I know a range of users, researchers and clinicians alike, who will be very happy to hear that. 😉

[thijsdhollander]

A short question about the data maybe: you mention "31 b1 volumes" but also "96 directions". I'm slightly confused, as the number of b1 volumes is typically what's "also" called the number of gradient directions. Do you mean the protocol has a few repetitions of this, for example 3 repetitions of a 1 b=0 and 31 b=1000 protocol?

To follow up on this bit, for future reference: I've clarified this together with @atefbadji, and it is effectively 31 gradient directions (and a single b=0 image), without any other repetitions. There was a slight confusion due to the number of numbers / lines in the bvecs file.

So this looks really good then: just a single b=0 image, 31 gradient directions, b=1000, anisotropic voxels; and everything still works very well. 👍 👍