Advice on using CSF signal for fMRI denoising

MRI Processing
1

I would like to ask the community for advise on denoising spinal cord fMRI data. Currently we use PNM with respiratory and cardiac recordings and use the resulting regressors in the FEAT analysis. Many talks and papers also suggest including signal from CSF as a regressor. Would you recommend it?

How much of the CSF should we include in the mask? I suspect CSF pulsation would be quite different depending on the vertebral level. Would it be best to only include pixels around the area we expect to see activation in?

Thank you,
Eugene

2

Dear Eugene,

Thank you for reaching out.

As suggested in Brooks et al. (2008), I would indeed recoment removing the CSF signal. It is pretty standard to add CSF as slicewise nuisance regressor (slicewise since it will indeed be different across the superior/inferior axis of the cord), and the first principal components (~5) of the signal could be used to get the voxels that show high variance. In this case, the size of the mask has less importance. Here are a few examples:

Let me know if you have any further questions,

Best,

Sandrine

3

Thank you, this is very helpful!

4

Dear Eugene,
I agree with Sandrine’s comment that it’s common practice to use the CompCor approach for fMRI denoising at both brain and spinal cord levels (e.g. 5 first components: ): see Behzadi et al. 2007

Here is an exemple to implement CompCor with tapas toolbox (matlab) for volume wise denoising. This approach was used for the following spinal cord papers: 1, 2, 3, 4

You can also use Nipype (python) to implement CompCor approach

Ideally, I recommend extracting the components slice-wise. For instance, you could create a loop to extract the PCA for each slice within your CSF mask (in individual space). This approach will accounts for signal fluctuations across different slices.

I would not use an a priori approach to select the portion of your mask. Instead, I recommend denoising by volume (using the entire CSF mask) or, even better, performing slice-wise denoising (csf confounds will be specific to each slice).

Best,

Caroline

5

Thank you, Caroline! We will try to implement a similar approach in our pipeline.