Hi Rick.

I take your point! I don't usually smooth time series in the volume so haven't given this much thought.

Two notions come up.. It seems to me that people who smooth time series in the surface domain might not run into this difficulty. Specifically, it seems to me that a workflow in which time series are projected to the surface and those are smoothed there won't run the risk to bringing in data from outside the brain. sumavol2surf will ignore voxels outside the cortex (assuming good alignment between functional and anatomical), and the subsequent smoothing will of course only consider vertices on the surface. In any case it's an interesting issue and i'm tempted to run a few tests to see how our noise behaves.

Second, one way to ignore voxels with low snr (low mean, high var) might be to derive an SNR mask and use the -1filter_blur option in combination with that mask to include in the smoothing only voxels that pass the SNR threshold. That will probably get rid of the potential influence of low SNR voxels anywhere.

Thanks,

Oori