At present, this capability does not exist, nor is it trivial to add. The reason is that the structure of 3dDeconvolve is to apply the same regression model to each voxel. Voxel-dependent censoring would mean a different type of equation would have to be solved for each voxel. I have thought about adding this capability in the context of prewhitening (which also changes the equation for each voxel in a different way), but so far this is only dreaming.
bob cox