Hi -

I’m trying to develop a better sense of how to interpret full_mask+tlrc (one of the products of the ‘mask’ block generated by in afni_proc.py) If you look at the chunk of code where this gets produced, you see a call to 3dMean that’s operating on the products of applying 3dAutomask to the [blurred] EPI data. The dox for 3dAutomask say (in part):

Method:
 + Uses 3dClipLevel algorithm to find clipping level.
 + Keeps only the largest connected component of the
   supra-threshold voxels, after an erosion/dilation step.

3dClipLevel dox say (in part):

Algorithm:
  (a) Set some initial clip value using wizardry (AKA 'variance').
  (b) Find the median of all positive values >= clip value.
  (c) Set the clip value to 0.50 of this median.
  (d) Loop back to (b) until the clip value doesn't change.
This method was made up out of nothing, based on histogram gazing.

So the idea is to use the variation in the signal to seperate ‘live’ or ‘signal-bearing’ voxels from background voxels. (This makes sense for EPI data; not sure how it works for anatomicals.) My assumption then is that voxels that do _not_ make the cut in full_mask are voxels that seem to be worthless in terms of signal-bearing-ness. For instance, here’s the mask for one of my subjects, overlaid on anatomy:

[www.dropbox.com]

We can see incomplete coverage in parts of temporal lobe where we would expect poor SNR due to signal dropout, so that makes sense; although the extent of the bare regions in the the LH of this mask is a little surprising.

My questions are:

0) Is my interpretation of what full_mask is, and what it represents, correct?

1) If my interpretation is correct, why are the full stats performed on data (the pb04 series) that has not been filtered by this mask? Is the idea that those voxels will be mostly noise, so the statistical testing will see that they’re worthless and not report anything there, so it all works out in the end anyway?

2) If that’s so, why is the signal-to-noise data (the TSNR dataset) gated by full_mask? The voxels that do not survive the multiplicative operation in 3dcalc using full_mask are just the voxels with very low SNR, right? So why not show them? Or are they inappropriate for inclusion for some other (non SNR) reason?

I looked around the msg board for similar questions and could find none. Any intuition-building insights would be appreciated.