Hi all,
I have a question about denoising. We're doing a task-based functional connectivity analysis. Our first-pass results showed patterns of activation in line with our hypotheses, but we were also getting a lot of artifact (e.g., white matter activation, and the classic "ring" pattern around the brain). This was even when we included WM, CSF, and motion regressors. Anyway, we thought we'd give ICA denoising a shot and indeed were able to get rid of many components that were clearly structured noise. We then re-ran our original analyses, and, comfortingly, got similar patterns of activation. However, the motion-related artifact is actually worse and we do lose a couple of regions that we had a strong a priori hypothesis would be activated.
Do you have any recommendations? Clearly our component selection was slightly off (either too strict or too lax). Since this was our first shot, we were fairly stringent in what we considered noise (>90% of the activation had to be outside the brain, WM, or periphery). We could probably add in a lot of additional noise criteria in line with Kelly et al 2012 or even use the training sets provided with the FIX plug in.
I guess what I'm having trouble with is understanding how removing noise could actually have increased the presence of artifact. Our selection of components was not entirely clear-cut, as there were many "borderline" components (for example, maps containing a combination of plausible activation and a motion-related ring). I've found some great resources on component selection on-line, but any additional insights on how to handle these "iffy" components would be very much appreciated. I'm worried about removing meaningful activation but I'm also unhappy with artifact present in my result maps.
Thank you,
Lizz
|