This output is a little bit of a complicated one. It is to do with obtaining
a tradeoff so that the warping does not overfit the data. Nonlinear
registration algorithms make a tradeoff between between keeping the warps
smooth (regularisation) and reducing the difference between the images. In
SPM spatial normalisation, if the fit is poor, then the registration is
regularised more heavily.
The measure used for assessing this tradeoff is based on the mean squared
difference between the images, but with a correction for the smoothness of
the residuals. If the residual differences are more smooth, then there are
essentially fewer degrees of freedom. This Var measure is the sum of squared
residual difference divided by the "degrees of freedom". These dof are the
number of sampled pixels, minus the number of model parameters all multiplied
by a smoothness fudge factor.
It isn't the best way of doing it because I didn't know about REML stuff back
then, but it kind of works.
Best regards,
-John
> I notice that the 'Var' in the output of spatial normalization is kind of
> an converging index.
>
> For example,
> iteration 1: FWHM=8.663 Var=1.96911
> iteration 2: FWHM=7.168 Var=1.05235
> iteration 3: FWHM=6.71 Var=0.914537
> ...
> ...
>
> Using a Var vs. iteration plot, it seems to show some convergence. But,
> what is the definition of this 'Var'? Secondly, what is the definition of
> the FWHM? (I know it's full width at half maximum, but ...). I am needing
> this information for explaining how SPM normalization works to my
> colleagues.
|