Hi Alfonso,

I read many of your posts in which you mention that CONN uses CompCor for the denoising step. I was wondering what exactly is CompCor doing, does it simply regress out the WM and CSF, or does it refer to the whole regression of confounds we list in the denoising step (eg WM, CSF, motion regressors)? I suppose band-pass filter and detrending are not part of the CompCor algorithm?
My understanding is that CompCor generates principal components (PCs) that can be used as covariates in a GLM. Is this automatically done in CONN, where can we find these PCs? And how can we use CompCor information when analyzing resting state connectivity (without a GLM)?
Thank you,
Ben

Hi Ben,

Here is what I think goes on under the hood (Dr. Alfonso can correct and/or add details):
The gray matter signal is band pass filtered and detrended. After that 16 principal components are extracted from the white matter and the CSF regions (these regions are saved as ROI files during the segmentation step; actually the files are eroded a few times before saving as ROIs). These, along with the six motion parameters, their first order derivatives, and other effects that you select in the denoising tab (for example, the main effect of the task), are regressed out from the signal of interest. You have the choice of deciding whether band pass and detrending happens before regression or simultaneously in the denoising tab.

These steps are performed before the first level modeling (i.e. the first level GLM is applied to already denoised data). Therefore, the same strategy can be applied to resting state: the GLM will have a single condition (called rest) in the first level.

I mentioned that Conn extracts 16 components; however, the default number of components that are regressed out are 5 (both for WM and CSF).

These components are saved here: /data/ROI_Subjectxxx_Sessionxxx.mat file. Once you load this variable into MATLAB, you will find a variable named data. The first cell has the overall GM signal while the second and the third one have the PCs from WM and CSF (you will notice that they have 16 columns). The rest of the cells correspond to signal (before denoising) across the ROIs that you have specified. The names variable in your workspace will tell you which cell corresponds to which ROI.


Hope that helps


Best
Pravesh

Thanks Pravesh, that was very helpful. However I'm still not sure if the CompCor algorithm refers to all the steps you just mentioned (BP filtering, detrending, regressing out WM, CSF, movement parameters, etc), or if it just refers to regressing out WM and CSF, and all the other steps are done on top of CompCor.
Also, CONN can regress out the main effect of task, which in my case is "rest". But how is this regressor calculated? I noticed that the regressor is convolved with the hrf, but I compared the values with spm_hrf and they don't match.
Finally the realignment parameters do not show up in the "all effects" list and there is no way to select them. They appear in the Covariates (1st Level) in the Setup tab though.
Thanks for your help.
Best,
Ben