SS3T-CSD on b=1000, 35 gradients directions, one b0.

[stellasm]

I have single shell (one b0 and 35 b1 volumens) data with low b-value (b1=1000). I applied common pre-processing steps (denoising, unringing, preproc and bias correction). I estimated the response function of each tissue through dhollander algorithm. Then, I performed dwi2fod with msmt_csd option and ss3t_csd_beta1:

• MSMT-CSD tissue compartment images (WM, CSF):
  
  

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

• MMST-CSD FOD images:
  
  

• SS3T FOD images:
  
  

Finally, I performed probabilistic tractography using the SS3T setting (I used the crop to slab option and a thickness of 2 mm at mrview):

• cutoff = 0.08:
  
  

• cutoff = 0.09:
  
  

• cutoff = 0.10:
  
  

Cheers,
Stella.

[thijsdhollander]

Hi Stella,

Thanks so much for posting an overview of your results and experience with the SS3T-CSD algorithm! This is another good example of very challenged ("clinical") data quality, in line with #3, #4 and #5. The result is thus also limited in quality, but nonetheless some GM-like (mostly cortical) signal does get filtered out. As with those other results I referred to, this is still surprisingly good for such quality data. You're challenged on all fronts here: low b-value, low angular resolution and only a single b=0 image. Some of the other pieces of feedback suggest that the low b-value and low angular resolution can still be coped with pretty well. The single b=0 is theoretically speaking also not an issue, but in practice it's the one that risks becoming the "weakest link" of your data very easily. Any artefact, kind of noise and the smallest misalignment of that b=0 to the other dMRI data has an immediate potential impact on the result (just as well for 2-tissue MSMT-CSD by the way).

Overall, the GM-like signal is picked up reasonably well in the correct locations. The quality seems better at the posterior end of the brain compared to the anterior regions. Looking more closely at the individual SS3T-CSD tissue signal maps as well as the tractography results even in those more challenged anterior regions, and reasoning about what impact a very slight misalignment of b=0 relative to other dMRI data would have there, I'm pretty sure that's indeed explaining the relative quality of those anterior regions in the result versus the posterior ones in your data. Nonetheless, by also showing us the 2-tissue MSMT-CSD result for the same data, it nicely reveals the main benefit here: at the low b-value, the (cortical and other) GM has more signal than the WM; this becomes evident in the WM map from MSMT-CSD, where the cortex stands out brighter than the WM. This supports the desire to model GM-like signal in its own compartment: otherwise, you get large WM-like signal in the cortex, biasing not only estimation of the WM-like signal, but also hindering tractography (more specifically: hindering the possibilities of finding a decent cutoff to stop tracking appropriately). Even though the SS3T-CSD result may not look exactly "clean": the essential thing is that GM-like signal is filtered out as much as possible. At a low b-value, the contrasts in the data required to succeed at this goal are limited, so not all GM-like signal might be filtered out; but every bit that is, is a nice step forward in the right direction.

So in the end, your SS3T-CSD tractography results look very good actually; again especially in the posterior regions. Really neat and clean separation of WM and GM there, better than what I expected would be possible; so that's great to see! 😃 Finally, as to the cutoff value: this is always a bit of a qualitative assessment ("guess") to judge well, but I think I'd prefer 0.09 in your case here. It deals better with some false positives, in particular in the anterior regions, than 0.08 does. On the other hand, 0.10 seems a bit too constrained: I'm starting to see some slight gaps within the WM. This might of course be due to a limited number of (total) streamlines, but in any case, I wouldn't take any risks. It's always a trade-off between false positives and false negatives (and by default, whole-brain tractography will have a load of false positives anyway), but I'd rather not risk missing structures, especially if you'd eventually move toward targeted tractography, in case of specific hypotheses or applications.

Thanks once more for the amazingly well documented feedback. I'm surprised (but then again also not surprised) by the amount of feedback using low b-value "clinical" datasets. It's certainly not the optimal scenario, but it shows there's lots of these data around. At the same time, it shows the method can actually still work well on these data, and allows to deal with one of the otherwise most annoying aspects of the low b-value data: the relatively high GM-like signal (relative to WM-like signal). Nice!

Thanks again,
Thijs