I am running a fixel-based analysis with multi-site data, and wanted to know if there is more robust way to control for differences in scanners than including scanner-specific column regressors in the design matrix. I know harmonization techniques such as COMBAT have been successful on DTI scalar images, but given that FOD images are 4-dimensional, I doubt application would be as straightforward. Perhaps it would be more appropriate to control for site differences more upstream at the level of fixel metrics, but that would involve manipulating fixelcfestats sourcecode in probably non-trivial ways. For now, I am happy just including the column regressors, but happy to hear other thoughts.
1 and 2 make sense, but I am afraid I am not sure how to implement suggestion 3. Would this be done by specifying some kind of ANOVA contrast for fixelfcestats?
For suggestion 1, I was under the impression that the same response function for everyone was needed for between-subject comparison to be valid (at the trade off of a less “true” FOD image since the responses aren’t subject specific). Will having different responses still allow for valid between-subject inferences?
Yes, I believe this is supported by the -variance option of fixelcfestats.
In general, the recommendation is to keep the response constant between all scans in a group study, because otherwise the responses might (partly) “absorb” the microstructural differences between the groups, leaving no (or smaller) differences to be found between the FODs.
However, it does not make sense to use the same responses across datasets from multiple sites (or same scanner with different settings) because the same amount of perfectly healthy WM will produce a different MRI signal, and thus a different response, on each site. In that case, the recommendation is to keep the response constant between all scans from the same site, but to use a unique one for each site. That way the site-specific responses will “absorb” the differences in scanner and scanner settings and the resulting FODs will be more or less comparable between sites. However, as this “correction” will not be perfect, you should still use suggestion 2.
a different average response for each site to account for differences in relaxation, diffusion weighting, etc as much as possible
Important to distinguish between different sites with the same protocol, and different sites with different protocols. The latter case absolutely necessitates separate response functions (as would indeed also be the case for different protocols on the same site). But if it’s precisely the same protocol, and one accounts for any residual global intensity scaling, one could argue either way. I don’t yet have a feel of cross-scanner bias data.
variance blocks to account for variance differences between sites
Would this be done by specifying some kind of ANOVA contrast for fixelfcestats ?
ANOVA is a class of hypothesis tests that can be applied to estimated parameters of a model. Variance blocks apply at the point of fitting the model to the data to derive those estimated parameters. So two very different things that shouldn’t be mixed up. Explained best in the GLM master document.