Hello,
I want to calculate a dense connectome from rfMRi: i.e. correlate all voxels pairwise to create a symmetrical matrix of correlation coefficients, and then threshold that matrix to distinguish significant correlations.
Naturally it's a huge matrix (even just the upper quadrant), which presents some technical challenges, but my main conceptual problem is that I don't know how to threshold it. Can anyone direct me to best practice here?
Many thanks,
Tom
Dear Tom,
that's a tough problem. Intuitively, I would not do this voxels-wise. There are some atlasses available that are thought to represent a brain parcellation based on rs-fMRI. I remember the AICHA atlas and the work by Gordon et al (2016, Cereb Cortex), but afaik there are even more atlases out there. With a ROI approach, your problem would be less severe, as you only end up with a matrix with a size of maybe 200x200.
Still, the issue with significance testing remains. The main problem is that in rs-fMRI you look at correlations, and its significance largely depends on sample size, i.e. if you just grab a large sample of >200 subjects from one of the public data bases, you might literally end up with all regions in the brain correlating significantly with each other at massively low p-levels.
There are many studies in the field that dodge this bullet by using different p-values without proper explanation, somewhere ranging from p
You can start with looking at https://www.biorxiv.org/content/biorxiv/early/2019/11/07/741595.full.pdf?%3Fcollection=
to see that it's not a settled issue even for parcel-based connectome. Not sure though whether it's realistic to use all those regularisations to the dense connectome
Thanks guys (both).
Christoph: I suppose the permutation thresholding approach would be to just make N independent time series, correlate them all pairwise, and record the statistics of the null distribution that way. I have the lengths of the series, which are already group averaged and processed, but I'm not sure how to permute them sensibly. I see that a parcellation-based approach will make things a little easier or more permissive, but I'd rather keep it voxel-level if possible; I'm building maps rather than predicting, so the significance factor is unfortunately necessary.
On another note, I saw you wrote a manuscript but didn't have time to review it; hope it gets in! Lesion-model-interpretation stuff. Do send it to me when you're happy with it, or let me know when it comes out?
Slava: Wow, that abstract is a frightening read. This is a good reference though, so I'll see how far I can get both by reading it and by looking at their references.
So this is more complicated than I hoped, unfortunately. I'll think some more on it then.
- Badges
- Science topic