> I have a newbie question regarding smooth and realignment: By default,
> SPM will smooth the image before the realignment, I am wondering what
> is rationale behind it, could someone recommend some references?
Various reasons - most of them relating to spead.
1) It reduces interpolation artifacts for low-degree interpolation methods.
The alternative would be to use much higher degree interpolation (slow), or
live with the consequence that images translated by half a voxel are smoother
than those translated by a whole number of voxels. The extra smoothness at
half voxel translations results in periodic dips in the objective function,
which can lead to suboptimal solutions.
2) Not every voxel is necessarily used when doing the registration, which is
suboptimal. Smoothing partially compensates for this by combining signal
from neighbouring voxels, while reducing the noise.
3) The optimisation uses a Gauss-Newton strategy. Smoothing makes the Taylor
series approximations more accurate, leading to faster convergence. It
probably also leads to greater stability (because a Levenberg-Marquardt
strategy for regularising the updates is not included).
There are many reasons for using less smoothing though. One of these is that
smoothing makes the algorithm more likely to interpret activations as motion.
As with many aspects of SPM, there are a number of tradeoffs to be made. If
I made the algorithm more accurate, then people would complain that it was
too slow. If I make it faster, then it is likely to be less accurate. This
is why users are able to customise various default options in order to make
it work in the most appropriate way for them.
Best regards,
-John
|