Dear MRTrix community (and probably especially @rsmith),
I’d appreciate any help regarding this connectomics analysis. My goal is to obtain structural connectomes for all HCP subjects and use them to identify the connections that correlate with some functional measures.
After normalizing the intensity of the DWI data for each subject in the HCP dataset, I ran whole-brain probabilistic tractography, followed by SIFT2 and tck2connectome with the Schaeffer200 atlas. To enable between-subject comparisons, I used the same response function (the average response function across subjects). Finally, I also summed the connectome values across the 17 Yeo networks.
Now, I ran PCA to identify the principal axes of variation across subjects as a way of reducing the dimensionality for subsequent analysis. I followed these steps:
- Multiply each subject’s connectome by mu (from SIFT2)
- Z-score each node across subjects (this step takes care of the variable scales in the connectome - some edges tend to be small and some large, and I want all of them to contribute equally to the PCA analysis).
I noticed that PC1 is strongly correlated with mu:
This seems to go against the very idea of using mu in the first place, as a way of standardizing the values across subjects.
I also tried a divisive normalization approach, where I divide each row in the connectome by sum of all connectome valuesright before running the Z-score and PCA. This results in a negative correlation between PC1 and mu:
This makes me wonder what the correct approach would be. Wouldn’t you expect negligible correlations between mu and the first PC? That is, that between-subject differences in the connectome would not stem from differences in the response function.
Many (many) thanks for your help,
Roey