| The brain was shrunken within the volume - the dimensions were exactly the
| same. I think your second suggestion is correct - normalizing with bilinear
| interpolation does not result in the same shrinking. Your answer would seem to
| suggest that the brains are not really shrunken, but rather that the outer
| edges are eroded by the interpolation procedure (leaving the internal voxels
| accurately registered) - do you think that's correct?
|
| does this problem argue that zeros, rather than NaN's, should be inserted into
| the masked-out voxels in the next version of SPM99 (or alternatively, that the
| sinc interpolation procedure be made more robust to NaN's)?
Yes, the resulting images would appear to be eroded at the edges. We have
thought about how best to handle the NaN regions, but since the problem
only arises occasionally (when you resample floating point volumes that contain
NaNs), I have not done anything about it. It could be partially fixed in the
interpolation functions, but it would require quite a lot of additional
computational overhead. You should run into fewer problems if you stick to
trilinear interpolation for resampling the con_* images.
All the best,
-John
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|