Manually filter streamlines before SIFT2


#1

Dear experts,

I am trying to investigate auditory pathway from inferior colliculus to auditory cortex. My data are b=1100, b=2500, 64 directions for each shell. I am running:

  1. tckgen with whole-brain seeding:
    tckgen WM_FODs.mif.gz 30M.tck -act 5tt.mif -backtrack -seed_dynamic WM_FODs.mif.gz -select 30M -cutoff 0.06
  2. tckgen -seed_image from mask of colliculi and from auditory cortex, 40 millions
  3. combining streamlines together
  4. running SIFT2
  5. counting weights using tck2connectome

So far I am not very satisfied with the result: the weights sums of pathway of interest vary very much between participants (some have around 10, some several hundreds) and are very individually left-right asymmetric. Therefore I am looking for a way how to improve tract identification.
I see that some streamlines follow non-plausible pathway through anterior pons. I would like to filter out these streamlines. Please confirm that following is reasonable:
define ROI in anterior pons
between step 3 and 4 (i.e. before SIFT2) run tckedit -exclude ROI_anterior_pons.mif
Would not this step violate rationale behind SIFT/introduce any unwanted bias to the results?

Antonin


Null Distributions
#2

So far I am not very satisfied with the result: the weights sums of pathway of interest vary very much between participants (some have around 10, some several hundreds) and are very individually left-right asymmetric.

Since the concept of combining whole-brain tracking with targeted tracking and then using SIFT2 has only been shown as a proof-of-concept, I would be curious to know whether the more “standard” approach (just extracting tracks from the whole-brain tractogram post-SIFT2) has the same magnitude of variability. The regularisation in SIFT2 kind of “resists” the combination approach (since you need very small weights for so many streamlines in your pathway of interest but weights ~ 1.0 for other streamlines traversing the same fixels), and so may not be effective in practise.

Therefore I am looking for a way how to improve tract identification.

Aren’t we all :sweat:

The use of streamlines tractography in subject space for quantification is inevitably going to introduce a lot of unwanted variability. The primary question I would ask from an experimental design perspective is whether or not FBA can be used to assess the pathway of interest, as opposed to using tractography / SIFT to assess the connection of interest.

I see that some streamlines follow non-plausible pathway through anterior pons. I would like to filter out these streamlines.

Would not this step violate rationale behind SIFT/introduce any unwanted bias to the results?

It’s reasonable to remove streamlines from a whole-brain tractogram that are considered false positives due to some other criteria prior to SIFT. What is not permissible is removal of entire macroscopic bundles prior to SIFT. Fundamentally, it must be reasonable that the FOD field can be fully explained by those streamlines present in the tractogram. If only a subset of the fibres contributing to the diffusion signal in a fixel are present within the tractogram, then SIFT/SIFT2 will erroneously boost the contributions of those streamlines that do traverse that fixel, in an attempt to explain the magnitude of the diffusion signal.

Rob


#3

I did a test on small subset (approx 10 subjects) and coefficients of variation of sum of weights in both standard whole-brain post-SIFT2 and combined whole-braint + targetted post-SIFT2 approaches are similarly bad = about 100%.

I am quite confused here. Not sure I understand the terminologic difference (pathway vs connection) but mainly the offer of FBA for this study case what is confusing me.

My understanding was that if there is a hypothesis on a connectivity of specific pathway which can be defined by starting and ending gray matter mask(s) (which I think is my case), then individual tracking, SIFT2 and sum of the weights of the streamlines connecting nodes of interest is always the method of choice.

In contrast, FBA is method of choice when there is not pathway-specific hypothesis and one wants to test whole-brain fixel-wise. Maybe also in the (rare) cases when in some pathologies only part of the pathway along its length is affected, or some macroscopic fibre bundle containing many different pathways is of interest, one can think of restriction of FBA to specific set of streamlines, as discussed in this post. However, also in this post, a question was raised whether single scalar measure is not a better option.

So do you see any benefit in doing FBA in case where hypothesis is concerning a connectivity of specific pathway which can be defined by gray-matter nodes the pathway involves? Do you think that one could gain a better sensitivity using FBA (with restriction of the fixel mask to the pathway of interest) when assessing a less identifiable pathway by avoiding the individual subject-wise noisy tracking? In FBA, the tracking would be performed only once and on less noisy FOD template. This approach would probably lower specificity to attribute the potentially observed effect to the particular pathway, but could it enhance statistical sensitivity to observe any effect in comparison to subject-wise tracking, SIFT2 and extracting edge-wise sum of streamline weights?

Antonin


#4

Not sure I understand the terminologic difference (pathway vs connection) but mainly the offer of FBA for this study case what is confusing me.

Well, I kind of made up that terminology on the spot, so the lack of understanding is not terribly surprising. :neutral_face:

Imagine two grey matter regions of interest, which are connected through a white matter “pipe”. However, there are also a bunch of other grey matter regions on either side of the pipe, for which the connections between them pass through the same pipe. If one were to assess the “pathway” (i.e. the pipe), using e.g. FBA, then the data you would observe would be a combination of both the connection you are interested in and a bunch of other connections. If the effect of interest is genuinely specifically just the “connection” between those two grey matter regions, then having a quantitative metric reflecting just that connection would be preferable.

This is however predicated on the combination of tracking & SIFT being able to reflect that underlying specific difference within the reconstruction; which may unfortunately be too much to ask in many circumstances. So while e.g. SIFT2 sum of weights is physically and logically the correct way to assess the hypothesis of an altered “connection”, it’s not yet clear whether this specificity is outweighed by the data variance introduced by the necessity to perform individual subject tractography.

In contrast, FBA is method of choice when there is not pathway-specific hypothesis and one wants to test whole-brain fixel-wise.

Typically yes; but “FBA” shouldn’t necessarily be equated with “whole-brain”, and indeed fixelcfestats now has a -mask option to specifically not do whole-brain analysis. What I was trying to comment on here was not so much the difference of targeted hypothesis v.s. whole-brain analysis, but more the underlying quantities being tested: intra-cellular volumes within fixels v.s. intra-cellular cross-sectional areas ascribed to streamlines connecting grey matter targets. They are very strongly related, but that relationship is achieved through the use of individual subject tractography, which comes with its own set of limitations.

Do you think that one could gain a better sensitivity using FBA (with restriction of the fixel mask to the pathway of interest) when assessing a less identifiable pathway by avoiding the individual subject-wise noisy tracking?

It’s possible. It would certainly be an interesting experiment to see through. However my point was that the hypothesis is not quite equivalent: The test would be of the fibre density values within fixels in which the pathway is thought to reside (and hence other fibres within the same fixels contribute to the result), versus quantifying the strength of the connection itself (which incurs the limitations of quantifying both the pathway via tracking and the strength via SIFT2).


#5

Sure! We’ve been doing this internally already for several studies where we do have such a hypothesis, and we may otherwise lack the statistical power for a whole brain FBA. Just go from targeted tracks (no SIFT needed) to a “fixel TDI” and threshold to get a decent (binary) fixel mask of the bundle of interest, and provide that to the -mask option at the stage of the stats. You can still (hope) to get fixel-specific differences, which may still be very informative for a longer tract, of which only a particular part may be affected.

Apart from that, you could just get the average FD across such a (binary) fixel mask, and use that single number / variable of interest. That would further boost your power, but decrease potential specificity. It may however be more sensitive to cases where the effect is not in a specific spot along the tract, yet may be in different spots in different subjects (yet still within the tract).

Take a look at this recent FBA we’ve performed: https://www.researchgate.net/publication/322339694_Fibre-specific_white_matter_reductions_in_Alzheimer’s_disease_and_mild_cognitive_impairment (published in Brain).

Here, we didn’t define tracts entirely manually, but we used the FBA outcome from one comparison (healthy controls versus Alzheimer’s disease), to be used as separate (tract-) fixel masks for an analysis of healthy controls versus mild cognitive impairment, where we otherwise lacked the power to get such results. So in this case, our hypothesis essentially came from another FBA (and we hypothesised MCI may be a precursor to AD in certain cases).