Two questions that are not entirely clear to me, please excuse me if they are redundant.
Do you recommend using ACT for tracking based on FODs generated by SS3T-CSD?
It seems to me, that nicely calculating WM-FODs by SS3T (Mrtrix3Tissue) already generates valid information on tissue type / tissue probability for tckgen (iFOD2).
As I understand, ACT also generates biologically realistic priors for gray matter and CSF (which are of course not directly present in SS3T-WMFOD). So that might be a reason to also use ACT.
if ACT is recommended for tracking with SS3T-FOD data, would it then be sensible to omit volume 4 of the 5tt map (“pathological tissue”)?
Specifically I am interested in (extensive) white matter lesions (WML of presumed vascular origin). Masking them out as pathological tissue would lead “no anatomical priors” being applied to the streamlines, therefore, it would solely rely on the underlying WM-FOD (with low amplitudes in WML due to contamination by free water / CSF-“Like” signal). what happens if I don’t mask them out? would this region then maybe classified as white matter and “override” WM-FOD amplitudes I so carfully generated with SS3T?
I am working with single-shell data (3T, 2x2x2 mm, b=0, b=1500, 62 directions)
Thank you for providing the software & support, it is much appreciated!
I hope you’re doing well! I’ve just noticed you contacted us a bit earlier about our stroke studies; I’ll get back to you on that as well. We’ve learned quite a bit in the meantime about lesions with (likely) underlying vascular origins. Happy to share some insights.
About your current questions: I think one of the main things to realise is that the choice of using ACT or not is relatively independent of having used SS3T-CSD earlier on in the pipeline. The only connection of sorts is the fact that “out of the box” tractography will look a bit nicer and appear already more constrained to regions where it should go (and less / not in those where it should stay away from) when using SS3T-CSD versus using only single-tissue or 2-tissue modelling. I’m guessing this property might be what had you wondering whether the anatomical constraints are “already taken care of”; but it’s a separate thing in principle. Whether to use it or not, likely depends more on what you want to pursue afterwards. So your questions aren’t redundant; although I think you came into this from a different angle maybe. I’ll address a few things just briefly:
So this depends on what you want to do with the resulting tractogram afterwards. If it’s e.g. connectomics, where it’s often important streamline endpoints hit the cortex, etc…, then typically yes, you want to use ACT. If it’s e.g. targeted tractography of specific structures, because you want to analyse particular bundles, etc… then often no, you don’t need ACT. In the latter scenario, you’re likely already yourself imposing the relevant anatomical constraints by means of your inclusion / exclusion / … regions for tractography. If that’s already doing the job, then you’re good. This is important to realise, as in the presence of lesions, you might actually face unique challenges in using ACT. This is a bit too much to write all down here, but some of it at least relates to other parts of your post, so read on below.
Well yes and no. These things are quite separate in some ways. Yes, it does make the WM FOD out of the box more specific to actual WM, and that’s a great thing for tractography. E.g. in cortical GM, the WM FOD will be much smaller, in line with reduced axonal density. The other GM signal from neuronal cell bodies, etc… ends up nicely “out of the way” in the GM compartment. So the presence of other tissues in the model improves the WM FOD, which greatly helps tractography. But it doesn’t perform an anatomical kind of “constraint” mechanism directly in the tractography process. So basically: it’s 2 different things. In lesions, we have used the SS3T-CSD approach as a diffusion signal representation rather than a biophysical model; e.g. for stroke you’re likely aware of our very recent work looking at lesion compositions. This is also great, because again WM FODs will decrease according to (e.g. reduced) presence of axon density, when infiltration of other tissues and fluids happens. But also, the WM FODs in all those regions will still retain nice angular structure because signal from other tissues is filtered out (and thus doesn’t “pollute” the WM FOD). Again, you’re set up with great WM FODs out of the box here as well, benefiting tractography but even other algorithms like registration, etc… greatly.
ACT depends on the quality of the segmentations (i.e. the 5TT map) that it uses. It then applies a set of rules to this, e.g. streamlines not allowed to stop in WM, streamlines need to end in / at GM, no traversal of CSF, etc… But herein lies the challenge: you need a good 5TT map. There are different ways to obtain this, but you should take a very close look to what your 5TT files look like out of the box in the lesions in your data. Because intensity-wise lesions are often in the range of GM or even CSF on a T1w image, several image segmentation algorithms might segment these as “GM” or “CSF” accordingly. That will then cause a problem in ACT if you don’t fix that up: if a lesion is labelled “GM”, ACT will treat it as GM, and thus make streamlines stop at the WM-lesion boundary. That’s likely the opposite of what you want. That’s why the pathology class exists. Again, read on.
Well, so likely you then need to actually use or create volume 4 (the 5th volume).
No, I think you have the wrong idea about what the volume 4 / pathological class does. So out of the box, you’ll likely have volumes 0-3, coming from e.g. 5ttgen. First have a look at all of those in regions where you have extensive WMHs present. As I mentioned above, I’m guessing maybe some WMHs show up in the “GM” segmentation. So just applying ACT out of the box like that, will cause the issue I mentioned above: the algorithm will apply the rules of actual (e.g. cortical) GM to lesions. However, you likely want it to be treated as WM, because it is WM. A WMH doesn’t make it non-WM. So you could either fix all of that up manually (or e.g. using some lesion segmentation tool to aid this) and remove those bits again from the GM image, and add them again to the WM image. That would fix everything in principle. If you were to not do that, but rather draw those lesions in volume 4, what happens during tractography is exactly this: in those particular regions, the ACT rules just get switched off. But this doesn’t affect the WM-FOD amplitudes at all, so you retain the full quality and amplitude of the FODs you obtained from SS3T-CSD. Also, the other tractography constraints and settings still get applied (e.g. your curvature threshold, etc…). So again, these things are mostly separate. However, you’d have to be extremely precise in segmenting those lesions, and maybe even dilate the segmentations a bit. If they’re even only a little bit “too small”, you’ll still have “fake” GM on the edges of lesions, and most of the problem still remains when streamlines hit that “fake” GM. Same thing for punctate lesions or e.g. other small de novo lesions. So it’s a bit of a nightmare to get the precision right! That’s why you have to carefully consider whether your application requires ACT.
No problems there; SS3T-CSD will likely work well. On the ACT side, 5ttgen typically works with a T1w image. Finally, you certainly need a means to correct for EPI distortions if you need to perform ACT using a 5TT image derived from a T1w image.
I hope that helps. I’ll get back to you on the email; those questions are a bit more complicated.
Thank you for your reply and extensive explanations. That helped a lot to clarify my questions.
Indeed I thought that SS3T-FODs (GM / WM / CSF) could be used for anatomical-constraint tractography similiar to using 5TT-maps. They capture different properties of the MR signal (representation of diffusion signal vs. T1-signal intensities). However, I thought that you could “inform” or “improve” your 5tt segmenations (based on T1w-data) by adding (for example) information from the SS3T-GM-FOD to get an estimate how likely your 5TT-GM-segmentation would indeed be GM-“like” in the SS3T-GM-FOD. But reading your recent publications again I understand that this might not be feasible / an oversimplified approach.
Regarding the 5TT map I made the same experience. One has to be extremely carful in patients at risk for cerebrovascular diseases when relying on T1-data for tissue segmenation (presence of lesions from acute stroke, older stroke lesions, WMH, microbleeds etc.). We are working with 1x1x1mm T1 data and commonly use freesurfer (aseg.mgz) as an input to 5ttgen. In our T1w-data, chronic stroke lesions (in deep white matter) are commonly misclassified as CSF or gray matter, whereas WMH is classified as “WM-hypointensities” in freesurfer, but weiredly enough sometimes as GM as well (which leads to strange-looking fibers starting in the deep white matter). We are mostly caclulating whole-brain large-scale connectomes, so that’s why we use ACT in the first place.
Concering WMH in particular, we are very happy with segmentation results generated by BIANCA (Griffanti et al., Neuroimage 2016), even with relatively small training dataset. Thank you for pointing out that these masks might need to be dilated a little to remove erroneous GM areas.
Theoretically it’s not unreasonable to do something like this. But arguably a T1w image will have the higher spatial resolution, which itself is a desirably property to “bring into” the process. Even if you do construct something along the likes of this idea, you’d still need a separate lesion segmentation or something at least to define them (and on that front, a FLAIR image will likely be the more valuable source to bring this information into the process): because of the very reason we use SS3T-CSD as a diffusion signal representation, you can have things like lesions appear as a mixture of some remaining WM-like signal, but with a presence of GM-like and CSF-like signals (with contextualised interpretations). In other words, each modality brings its own strength.
Yep, not surprised at all. It makes sense, given the information these segmentation strategies have. Some do better if they make some assumptions regarding location and shape or continuity of known structures, but there will always be scenarios where that isn’t enough. And then you get the things you observe here as well.
So that makes sense. But if you’re after connectomes, all prerequisites to have a decent and aligned T1w image are likely already in place (you typically need these to get your parcellation anyway).
Yeah, we had good experiences with it too indeed. In terms of connectomes, key is indeed to make sure streamlines can’t “hit” false positive GM. Even a single voxel in the middle of dense WM can be slightly disastrous: lots of streamlines likely hit it. Just make sure to at the same time not wipe out (too much) of genuine cortical GM. And finally, of course, be prepared to deal with a decent amount of variance in the end.