Good Morning Community,
I have started working with MRTrix some months ago and I found it very useful for my research.
However, I need an opinion about the fixes-based analysis. My Boss gave me some data to run this kind of analysis. The only issue is that the data are from two different scanners and have different b values. One has 1 b0 and 30 b=2500. The other one has one b=0 and 46 b=1000.
I know that, with DTI metrics, often people used the scanner as covariate or used harmonization procedures. However, I am not sure if I can analyze these data together (e.g. with scanner as covariate) with the fixel-based analysis. I think that this analysis is very related to the b values, correct?
I hope that someone can help me,
I am not up to date with the state of the art in data harmonization. I’ve seen publications that have used it and produced reasonable results, but there’s a little gremlin in the back of my head that remains sceptical. If it were me personally—and if the data supported it—I’d be dealing with the data differences during processing and at the statistical inference step.
In reality this is an exemplar case that I really should be writing up as part of an FBA wiki at some point. But here’s the basic points (and just my opinion, not a guaranteed best solution):
- Derive two group average response function groups, one for each acquisition scheme.
- Following FOD estimation, combine all data together for analysis.
- It may be preferable to use an equal number of participants from each acquisition for template construction to minimise bias against acquisition scheme.
- At the statistical inference step:
- Use scanner as a nuisance regressor.
- Use variance blocks to estimate separate variances for the data arising from the two schemes.
For this to work, there needs to not be a strong correlation between acquisition and your effect of interest. E.g. If you were interested in the difference between group X and group Y, but group X were all acquired with b=2500 and group Y were all acquired with b=1000, then it’s impossible to discriminate between group difference and effect of acquisition. If however you had some group X’s and some group Y’s acquired with b=2500, and same for b=1000, then the model can separate out those two different effects.
I will follow your suggestions.
Also, I have both groups acquired in both scanners, therefore, I can use the scanner as nuisance regressor.