Hi,
I do not completely understand your reasoning but yes: it is legitimate to identify IC based on their spatial appearance and to remove these but this may lead to false-negative results for the analysis of the filtered data. It is also true that external RF interference usually shows up as stripes in the EPI images. However, if melodic catches these stripes they are obviously not Gaussian, and as you describe their spatial structure clearly isn't! The effects of filtering them out by melodic (i.e. changes in the effective df, possible increase in FN etc.) will depend on the time-course characteristics and the number of ICs. In general, I would not expect RF interferences to be temporally correlated with the paradigm but depending on your setup their on/offset may (which would be bad because the filtering simply regresses the ICs out). And if you have just 100 time points and 20 ICs you want to filter out, for example, this would certainly be messy and not recommended.
Hope this helps-
Andreas
________________________________
Von: FSL - FMRIB's Software Library im Auftrag von Meredith Braskie
Gesendet: Fr 22.06.2007 01:20
An: [log in to unmask]
Betreff: [FSL] Gaussian noise in ICA
Hello FSL group,
I have a Melodic question for you. I was reading a post from last November discussing removing
components of noise. Christian mentioned that typically in denoising one removes components
based on structure, which makes sense to me. The post said that such a method includes removing
non-Gaussian noise, which if anything inflates the false negative rate. I have some noise in my data
that appears as striations across axial slices (possibly RF noise?) throughout several of my scans.
Components that are entirely made of these striations are almost completely Gaussian. Is it
appropriate to select components with these striations based on structure first, but then remove
them only if they are highly Gaussian (which happens unless there is something other than striations
occurring in that component)? I would like to remove the noise related to this artifact without
removing signal that may be related to the task, but I want to make sure that this will not increase
the false-positive rate because of removing some of the variance (as was discussed in the post). In
other words, does what was written still apply if the components are identified first structurally?
Thank you,
Meredith
|