Fixel-based analysis results

Hello,
I have run a fixel-based analysis to compare two groups of Alzheimer’s with and w/o depression. It is the first time for me, therefore, I am not sure about the results. Even if the acquisition is a ‘standard’ DTI single shell acquisition (b=0 and 64 directions with b=1000) I used dwi2fod with the option msmt_csd. Is it fine or is it better to use the ‘csd’ option?
However, from my results I had differences in several ares in FDC and in log FC (FWE<0.05). The results are very similar. On the other hand, I did not have any difference in FD (FWE>0.30). Is it possible this thing?
Thanks,
Andrea.

2 Likes

Hi Andrea,

We would generally advise using the msmt_csd algorithm even for single-shell data. The more pertinent question however is how you used that algorithm. If you provided it with three tissue response functions, but data containing only two unique b-values, that won’t work. With two unique b-values the algorithm can fit up to two tissues, in which case we generally recommend providing it with the WM and CSF response functions.

From your description / my experience with DWI sequences, I’ve a suspicion that you have one and only one b=0 volume. It’s likely still beneficial to make use of it (as opposed to using only the b=1000 data and only the WM response function for that shell), but it’s worth being aware that anywhere that that one volume has any kind of artifact is going to have a large influence on the resulting multi-tissue fit.

Lower b-value data are influenced by a wider range of factors than are high b-value data, so it’s not surprising if they struggle to find results in FD, though they’re certainly capable of doing to. I’ve not actually thought about whether low-b or high-b is more likely to have unwanted variance arising from inconsistency in fixel segmentation, which can hurt statistical power… You could have a look at this by choosing any random fix of interest, extracting the subject FD values for that fixel, plotting the values, and looking at the distribution. See also @sgenc’s manuscript on FBA at different b-values here. FC conversely is all about morphology so is actually reasonably invariant to b-value (as long as it’s “enough to resolve crossing fibres in the FOD template”).

Cheers
Rob

Hi Rob,
thanks a lot for your answer. Yes, I have only one b=0 and the acquisition is a clinical acquisition and the data are not very good. Maybe, it is better that I used the ‘csd’ option and see what kind of results I have.
Additionally, I would like to understand if there is any kind of ‘correlations’ between fixel values and connectometry. For example, if I have lower FDC/FD, I should have higher connectivity or not?
Thanks for the paper.
Andrea.

Maybe, it is better that I used the ‘csd’ option and see what kind of results I have.

You can still use the msmt_csd option, which has a hard non-negativity constraint instead of the soft constraint of the original implementation. But you would need to provide only the b=1000 data as input, and have a single response function that contains only one row.

Additionally, I would like to understand if there is any kind of ‘correlations’ between fixel values and connectometry. For example, if I have lower FDC/FD, I should have higher connectivity or not?

This to me looks like two different questions; I’m unsure if this is a mis-use of the term “connectometry”.

  • If you are referring to the method dubbed “Connectometry”, then yes, the two are related; indeed somewhere I have an entire manuscript draft comparing the two.
  • If a mis-use and you are intending to refer to endpoint-to-endpoint connection density, then depending on how you construct the latter, they may be related; see this manuscript for an explanation of methods that explicitly make them related.

You can still use the msmt_csd option, which has a hard non-negativity constraint instead of the soft constraint of the original implementation. But you would need to provide only the b =1000 data as input, and have a single response function that contains only one row.

Okay… Only to understand better, therefore, by using the dwi2fod with as input the DWI (with bval and bvec) without the b=0? Only the b=1000?

About the connectometry, I also create the connectivity matrices for all subjects, by using tcksift2. I found that in several brain areas I had higher connectivity for AD with depression than AD w/o depression, while in FBA analysis, I found inverted results and in different locations. It is for this reasons that I was asking if there are some relationships between these two analyses.

Andrea

About the connectometry, I also create the connectivity matrices for all subjects, by using tcksift2.

Okay, so you’re using the term “connectometry” to refer to streamlines-based endpoint-to-endpoint connectivity. This is drastically different to a published method that is called “Connectometry” as linked above, so I’d probably advise avoiding use of the term in this way, since it’s likely to lead to confusion.

I found that in several brain areas I had higher connectivity for AD with depression than AD w/o depression, while in FBA analysis, I found inverted results and in different locations.

There’s been a number of reports of “unexpected results” with SIFT2-based connectivity. However there are just so many moving parts in such an analysis that I really don’t know what conclusions to draw. I do have an outstanding hypothesis regarding how the specific mechanism by which the iFOD2 tractography algorithm works could lead to seemingly paradoxical connectivity results, and various ideas for alternative tractography algorithms that wouldn’t possess such a bias, but have never gotten around to testing it.

Between streamlines-based connectivity and FBA results on the same data, I would personally place greater trust in results derived from the latter: the former’s reliance on individualised tractography reconstruction, which is known to be highly corrupted by false positives, makes me more sceptical. But conversely it’s difficult to draw much conclusion from a single such comparison, especially when the statistical significance is not only in the opposite direction but also in different locations, and also when the mechanism of false positive correction in the former has not been elucidated. If tractography were trustworthy, then I would expect there to be a reasonably strong correspondence between connectome-matrix-based inference on FBC and FBA inference on FDC; but SIFT2 can only correct for erroneous density fluctuations given a set of trajectories, it can’t fix the trajectories themselves if they are incorrect.

Cheers
Rob

Hi Rob,
Thank you so much for your all explanations. I have really appreciated them.
I will double-check better all my results.

Andrea