Hello,
I would like to ask about your experience with deriving mean signals from WM/CSF masks. In particular, I am following this procedure:
- Use a specific mask in 1x1x1 resolution and threshold it at p > 0.9.
- Erode the masks several times (approximately 4 times for WM and 3 times for CSF).
- Resample them to the fMRI data resolution (for example, 3x3x3).
- Derive the mean signal from each mask.
I am curious about the minimal acceptable number of voxels for these masks. While I understand this is subjective and depends on the data, I wonder if there are any general guidelines. I’ve experimented with various thresholding and erosion levels, and with higher p-values, the masks often become empty. With the setup I described, the signal looks well-cleaned, and there isn’t too much correlation with the mean signal from GM. However, for CSF, I sometimes get a mean voxel count as low as 40 (with a minimum of 13). I’ve seen papers where erosion is applied to the point where masks contain as few as 5 voxels, but I would appreciate your thoughts on this. I tend to prefer more aggressive erosion to avoid any potential signal contamination from GM.
For context, I’m working with rs-FMRI data with 152 timepoints.
Thank you for any input on this matter.
Best regards,
Szymon