Using mrclusterstats to analyze complexity

[rgrazioplene]

@Thijs

As Rob reasoned too, in principle this is not a problem for CSD: even though the data has lower contrast, the response function should represent this too. Deconvolution acts in a way to normalise for this. However, it’s harder at low b-value because you’re essentially trying to sharpen a smoother contrast more, to get the same sharp FOD. So expect more variance: noisier FODs, maybe spurious peaks (with consequences for that complexity metric, as Rob alluded to).

Thank you both! I am using a group average response function, so I think noise shouldn’t be a problem at the between-groups level.

@Thijs

As to FD at lower b-values, that also becomes slightly more problematic, as with lower b-value, there will still be more signal left from, e.g., free water. So as the b-value goes down, gradually the interpretation of FD breaks down. To relate this to my earlier example of, e.g., free water replacing myelin in a certain scenario: this may in fact in theory even cause FD to go slightly up. However, even for b=1000, I expect such effects (in the case of my free water example) to still be quite subtle. I don’t see them “overpower” any “real” differences in FD.

So long story short: I mostly expect you to potentially have less power due to possible increased variance, and more difficulties to correctly deal with the structure of FODs (while segmenting them or trying to match subsequent fixels in case of a fixel based analysis).

In hopes of adjusting (at least somewhat) for free water, I am currently redoing my analyses using the dhollander option in dwi2response (as suggested in another post of yours). Ah, clinical data. I’ll reply here if it makes any difference in my results!

Happy to hear any other suggestions for due diligence, interpretation, etc. regarding the low b-value data.