Hi all,
Has anyone used the null distributions (Nulldist1 or 2)? I am struggling to come up with a meaningful interpretation of these null tractograms. The best that I can manage is that they give you the “tracts” that would be produced in an isotropic brain. But how is this useful?
I was then thinking to look at particular statistics of these null streamlines like the length, termination etc… but I am bumping into the same conceptual problem.
Does anyone have any thoughts on the subject.
Claude
this thread and therein recommended paper might be helpful for you:
Morris, D. M.; Embleton, K. V. & Parker, G. J. Probabilistic fibre tracking: Differentiation of connections from chance events. NeuroImage, 2008, 42, 1329-1339
Thanks for pointing me to the thread and the article. I had read them both though I still have a lingering question about what the physical interpretation of a null streamline would be and was wondering if anyone around here had specific thoughts.
I’ve been wondering the same thing. My guess is that you could statistically compare streamlines generated by iFOD2 with those generated by NullDist2 (the most sensible might be streamline count). If they are significantly different then this is an indicator that the streamlines generated by iFOD2 are not just noise. Would love to hear other people’s thoughts though.
In terms of what these algorithms could be used for currently, I would very much limit oneself to the mechanisms described in the original manuscript cited by @Antonin_Skoch. The interpretation is pretty much what @jjmcfadyen describes, though in a slightly more statistically rigorous manner.
Importantly:
I am struggling to come up with a meaningful interpretation of these null tractograms.
These algorithms are intended to be used in conjunction with targeted tracking, not for producing “tractograms” (which generally implies whole-brain tracking).
As to what other nefarious activities could plausibly be done with such algorithms, and therefore why I went to the effort of implementing them despite the fact that I never do targeted tracking? … no comment
In the context of whole-brain tracking, could this method be used for effective edge-wise thresholding of connectivity matrix (to distinguish significant connections) for the purpose of generation of binary tractograms which are needed for example for graph analysis (small-worldness etc.)?
If one were interested in producing a binary connectome, then yes, that would indeed be an interesting and novel application, and probably more robust than the techniques being employed currently for that purpose. Personally I don’t have any interest in any analysis predicated on a binary connectome given quantitative tract tracing shows inter-areal fibre counts that vary by more than five orders of magnitude; but don’t let my stubbornness stop you
I’ve generated some data using NullDist2 and have come back to this thread with a few questions about what to do next:
Does it make sense to use the NullDist2 algorithm with ACT?
Does it make sense to run SIFT2 on streamlines generated by NullDist2?
To show that the likelihood of a pathway generated by iFOD2 is greater than chance, we want to find a significant difference between the streamline counts (or maybe the summed weights of SIFT2?) or the iFOD2 data and NullDist2 data. Does the direction of the difference matter? (i.e. what if there are more streamlines for the null distribution; does this mean the likelihood of the pathway is less than chance?)
… Did you install a listening device in my office? I was discussing this with a colleague about a week ago
Does it make sense to use the NullDist2 algorithm with ACT?
Absolutely.
However, based on question 2, I suspect the question you’re actually trying to ask (or maybe a missing question 1.5) is: “Does it make sense to use the NullDist2 algorithm for whole-brain fibre-tracking?”. This is more difficult; the method was published with the intended application being targeted tracking (note that it’s entirely possible, and recommended, to use ACT when doing targeted tracking). Applications of the method beyond the intended use are not covered under warranty.
Does it make sense to run SIFT2 on streamlines generated by NullDist2?
… Maybe … Appropriate construction and interpretation of such an experiment is left as an exercise for the reader.
Does the direction of the difference matter? (i.e. what if there are more streamlines for the null distribution; does this mean the likelihood of the pathway is less than chance?)
You’d have to follow the derivation of the statistics within the original publication to make sure that the logic & mathematics hold for the inverse case. I suspect it would, given it’s essentially a Z-score based on standard deviation drawn from a Poisson distribution. This is however beyond the scope of what is provided in tckgen.
I am now very interested to investigate whether the result of tractography in my study I described in earlier post ever makes sense
I was considering to repeat the same experiment (all steps 1-5 I described in the post linked above, possibly also with manual tract filtering using anterior pons ROI) with nulldist2 algorithm in tckgen to determine if the streamline weights sum of auditory pathway obtained by tracking with measured FODs are significantly higher than when using nulldist2.
When I investigate one pathway, is my experiment targeted tracking or whole-brain tracking? Does the theory under Morris et al. paper apply for this case?
Thinking of that in principle: With SIFT2 on streamlines generated by nulldist2, how can I properly generate my isotropic FOD needed for SIFT2? I think that its integral cannot be constant, it should be modulated by either white matter content within the voxel (from 5tt) or by total FOD integral of my reconstructed FOD.
I’m using the Null distribution algorithm in targeted tracking analyses.
As the tracking is random, I get a different different tck each time I run it.
Should I make multiple iterations to capture the different tcks (and make sure that the measures I’m obtaining are stable and won’t change in the next iteration)? Or is a single iteration enough?
The question you are posing is actually not specific to the null distribution tracking algorithms. This is entirely a question of streamline count. If the outcomes of tractography - regardless of algorithm, and however you are measuring that “outcome” - changes to an unacceptable magnitude if you re-run tractography, you’re not generating enough streamlines. Obviously the precise trajectory of every individual streamline will always vary if you re-run probabilistic tractography; it’s about whether or not the information you extract from the tractography experiment is adequately reproducible.
