 |
 |
> >The driving force in the registration is based on minimising the sum of > >squared difference between the voxels in the image and those in the > >template. However, this is based only on voxels that are present in both > >images, so if a region of the template head is not present in your image, > >then the spatial normalization will ignore this part of the template. > > > > This confused me a little bit. My experience with fMRI images was > different. When I tried to normalize a "slab" of the acquired brain to the > SPM template, I generally got an abnormally stretched brain. This wasn't > unexpected because I was trying to fit a partial brain to a whole brain > where the two brains have different sizes, intensities and shapes using a > non-linear least squares algorithm. This would easily diverge to an > abnormal shape. We overcame this problem by acquiring a whole brain EPI in > each study and performing spatial normalization on this and using the same > normalization parameters on the partial brain EPI images. I would expect > success in registering a slab of the brain, say from an EPI image, to the > whole brain T1 weighted image both acquired from the same person, but how > does SPM decide which parts of the template to ignore in spatial > normalization of a partial brain image? Brain is not a structure with > distinctive landmarks every few milimeters. SPM minimises the sum of squared differences between the image and template. During each iteration, it scans the voxels of the template image and finds the current estimate of where the corresponding voxel should be in the image that is being normalised (object image). If this voxel lies outside the field of view of the object image, then this difference is ignored. Otherwise, the program includes this pair of voxels in the matching procedure, and adjusts the parameter estimates in such a way that the sum of squares difference would be reduced. One of the differences between the spatial normalization of SPM96 and SPM99b is that SPM99b not only does not include voxels outside the FOV, but also excludes those that lie less than 8mm from the edge. The smoothing that is done prior to estimating the parameters causes edge like artifacts that have been responsible for some nasty warps. Regards, -John %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%