I have a resting state data set that was collected with a 32-channel coil using multi-band acceleration factor 4. The data seem excessively noisy; the output of @ss_review_basic shows that many participants have up to 90 percent of their volumes being excluded due to “motion” (most are around 50%), but visual inspection of the data does not confirm this for most participants. I’m using 3dpc (following example 11, but using an ANTs-based tissue class segmentation) to find the principal components on an eroded CSF mask, and using 3 components in the deconvolution via the ortvec option. So my first question is that it seems like this excessive noise is coming from the principal components not capturing all of the noise in the data resulting from the multi-band sequence. Instead, it seems like using 5 components may work better - does that seem reasonable? I don’t actually understand why multi-band would make the data noisier.
More importantly, in reviewing the demeaned motion files, it seems like 3dvolreg is picking up on the effects of the multi-band sequence and inappropriately labeling volumes as containing motion artifacts. Does that seem possible? If so, I’m worried that first calculating motion and then running a PCA is double-dipping some of the same sources of noise. Would it be appropriate to first run a PCA on the raw data (i.e. Resting+orig, before it is moved into template space and scaled) and then regress out those components before any motion calculations are conducted?