Fixel-based analysis - assign significant fixels to tracks

Dear experts,

after obtaining significant fixels in fixel-based analysis (FBA), is it possible to select and identify tracks to which the significant fixels belong? I think this could help much in interpretation of results of FBA but I did not find any information of this possibility.

Regards,

Antonin Skoch

Hi Antonin,
Not sure if you have seen this thread. This will allow you to crop the template tractogram to show only streamline segments associated with significant fixels. This can be useful for visualising results in 3d.

If you wanted to see the entire streamlines that extend from the significant fixels, we don’t currently have an easy way to do this. @rsmith it might good to modify tckedit in the future to allow streamline selection with fixel data.

Cheers,
Dave

Hi Dave,

yes, I know about this visualization option discussed in the thread you mentioned. It seems very helpful.

Actually I meant this option. I was not too much interested to visualize the streamlines corresponding to particular fixel, but to assign specific gray matter connection to it. This could enable to identify connectivity of which specific parcels is altered at the group level and would add great potential for interpretation.

As I think about this, I wonder how would be the actual accuracy and precision of such potential fixel-track assignment, given the tractography is performed in template space, only once per whole group. In this context, several questions raise in my mind:

What is amount of smoothness in common FOD template? Is the template-based FOD and subsequently tractogram sufficiently “sharp” to identify also more subtle connections?

What is the precision of subject-template registration? Given complex white matter structure, could we expect to assure correspondence between subjects and template for individual subtle tracks or could this assignment be valid only for main white-matter fascicles?

What do you think?

Antonin

Hi Antonin,
Both tough questions!

Factors such as your image resolution and number of subjects will influence the spatial smoothness of the template. In general, larger structures in the core of white matter should be reasonably well aligned regardless, but structures that are either small or near the periphery of the brain (where there is more anatomical variance across individuals) may not be as well aligned, and therefore not well defined in the template.

When considering template-based FOD tractography there is also additional angular smoothness. Variance in the FOD peak orientations (e.g. from noise, natural anatomical variance, or errors in spatial alignment) cause the (averaged) FODs in the template to be smoother (i.e. peaks broader). This can cause additional ‘spread’ in probabilistic tractography compared to tracking on an individual.

I would say mainly valid for major white matter fascicles. However, smaller structures such as the fornix, anterior commissure tend to align well too. You may want to load up several subjects in mrview then flick between them if you have a hypothesis about a particular structure. You could also try decreasing the spatial regularisation (smoothing parameters) in the registration to see if it improves things.

Either way, I think the alignment should be better than FA-based registration, and/or FA-based skeleton projection.

We have been discussing the option of supplying a MNI-space FOD template. This would be useful for assigning tracts to MNI-space GM parcellations or fMRI results.

While we don’t currently have software for easily extracting streamlines that pass through significant fixels, you could take a more manual approach of:

  1. obtaining a mask of significant fixels using fixelthreshold
  2. converting this to a voxel-wise mask using fixel2voxel
  3. extract streamlines that pass through voxels in the mask with tckedit -include. This will work well in voxels with single fixels. However, if your significant fixels are in voxels with multiple fixels, you could then manually remove any streamlines that are extracted and associated with non-significant fixels (in the same voxel) using manually drawn ROIs and tckedit -exclude.

You could always use tcknormalise to warp the resulting streamlines back to the space if an individual if you want to compare with GM parcellations. Alternatively you could also warp the significant voxel mask back to the space of the individual and perform tracking (and editing) there. However, GM assignment from doing this on just one subject may not generalise to all individuals.

Hope this helps,
Dave

Actually I meant this option. I was not too much interested to visualize the streamlines corresponding to particular fixel, but to assign specific gray matter connection to it. This could enable to identify connectivity of which specific parcels is altered at the group level and would add great potential for interpretation.

This can certainly be done, though I would advise caution. The statistical inference is performed at the template fixel level, and that is therefore where the interpretation should ideally be applied, particularly when only a subset of the relevant pathway is identified as significant. E.g. Imagine a single fixel being identified as significant, and then performing probabilistic tractography from that fixel with a high degree of orientation dispersion: A large number of grey matter - grey matter connections could conceivably be generated, but labelling all of those pathways as ‘altered’ would likely be erroneous.

Such interpretation is better reserved for analyses where the inference is in fact performed on the endpoint-endpoint connectivity directly

1 Like

Dear experts,

to revive this old thread, in the meantime a paper


was published, where

“significant fixels were categorized to 11 white matter tracts, using anatomical DTI atlases to guide categorization (Wakana et al., 2004; Mori
et al., 2005; Oishi et al., 2009; Zhang et al., 2010)”

Unfortunately, more information how precisely this was achieved, was not supplied.

Could you please elaborate to help me how to reproduce this methodology?

How the figure 6 was produced? I can imagine to do it as follows:

  1. register FOD template (its l=0 image) to MNI (using e.g. FMRIB FA template )
  2. transfer some ROI-based WM atlas to FOD template space

then, using Dave’s recommendation

and then multiplying this mask with the WM atlas masks?

Was it done this way?

This approach seems suboptimal to me, since it does not exploit study-specific WM-FOD-based tractogram. Better way I see to perform anatomically-informed tracking on WM-FOD template or ROI-based streamline identification on the WM-FOD-based tractogram, and using this info to segment fixels to major tracts. But I do not know how to do this in automated way. Maybe some adaptation of AutoPtx scripts could do the trick?

Could you please comment on?

Antonin

Hi @Antonin_Skoch,

This is quite an open-ended story actually; you could easily think of wildly varying ways to approach this, and I reckon there’s absolutely no one size fits all approach. It definitely depends on your exact goals. Just to bring up one (of many) aspects someone might consider or value more (or less!) is whether it needs to be some standardised solution, so results can be compared / “reproduced” between different independent studies. On the other hand, given FBA results, you might form more specific hypotheses, and might be after a very particular subdivision of results into structures (which might often be “tracts”, but don’t have to limited to such).

In any case, I know Remika did a lot of manual curation here. This made sense given a set of particular set of structures stood out relatively well in the FBA results (but again, this isn’t always the case for FBAs). It’s indeed wise to use tractography to aid this process. This might be template tractography, but also here you could think of different, and I reckon sometimes far more appropriate solutions.

Automation of any of these processes is another consideration; sometimes desired, but sometimes probably not advised. One important thing to take into account here, if you’re pursuing automation, is that most current whole brain tractograms (whether on an individual subject or a population template) have a lot of false positive individual streamlines, up to the point they’re easily dominated by large numbers of false positive streamlines bundled into coherent false positive tracts. What possibly worsens the whole situation, is that FBA results themselves might show a degree of bias towards these very false positive connections if the (“true”) results appear near certain challenging geometries which drive false positives in streamline tractography. I reckon some degree of curation, whether up front or driven by the study results themselves, is always desirable when it comes to relying on whole brain tractography results.

I’ve recently been advising and working with people effectively approaching this kind of challenge from different angles; only to reinforce my thinking about this being a very broad class of solutions for challenges with different requirements… Always glad to hear about your insights if you pursue a certain strategy; but I think, in conclusion, the key message from my end is that it depends on what your study/problem/… at hand calls for specifically. First define the problem, then design the solution. :wink:

In any case, I’ll grab @rmito’s attention to contribute some details here on how she approached this for the referenced work! I hope the above already inspired some fresh thinking on this topic!

Cheers,
Thijs

Hi @Antonin_Skoch,

I can give you some more specific details on what we did in the paper, but as @ThijsDhollander says, there are a number of approaches you can take!

Here are the specific steps that were taken in our case:

Firstly, as per Dave’s recommendation:

  1. Threshold significant fixels using mrthreshold
  2. Create signficant voxel mask using fixel2voxel

The next step required quite some manual curating:

  1. Manually create inclusion and exclusion ROIs to define regions where tracts will pass

Here, we referred to existing WM atlases. The referenced atlases were used to ‘guide’ categorisation of fixels into tracts in this way, rather than registering atlases to FOD template space.

  1. Create tracts-of-interest using tckedit on the FOD-based tractogram, with inclusion/exclusion ROIs, and using the significant voxel mask
  2. Generate fixel TDI of streamlines using tck2fixel
  3. Threshold TDI to generate fixel mask using mrthreshold

Following these steps, mean fixel-based metrics could be computed within each tract using mrstats

To address some other comments:

Fig. 6 in the paper shows the tracts-of-interest created in step 4. The tracts correspond only to the significant fixels, and fixels were grouped into tracts through categorization that involved manual ROI labelling.

The ROI-based streamline identification is essentially what was done here. I don’t think there are currently good automated options available, but if you find a solution, I’d love to know too!

Hope that helps!

Remika