One of the main sources of difficulty when comparing anatomical data from a number of subjects is anatomical variability. Even after automatic linear registration , there remains a residual non-linear component of spatial mis-registration among brains that is largely due to normal anatomical morphometric variability that can account for a difference in position of up to 1-2cm for the same anatomical landmark between subjects.
We have developed a completely automatic method, based on multi-scale, three dimensional (3D) cross-correlation, to non-linearly register two MRI volumes together, thus reducing the residual mis-registration remaining after linear alignment.
Spatial registration is completed automatically as a two step process. The first  accounts for the linear part of the transformation by using correlation between Gaussian-blurred features extracted from both volumes. In the second step, ANIMAL estimates the 3D deformation field  required to account for this variability. The deformation field is built by sequentially stepping through the target volume in a 3D grid pattern. At each grid-node i, the deformation vector required to achieve local registration between the two volumes is found by optimization of 3 translational parameters (txi,tyi,tzi) that maximize the objective function evaluated only in the neighbourhood region surrounding the node. The algorithm is applied iteratively in a multi-scale hierarchy, so that image blurring and grid size are reduced after each iteration, thus refining the fit.
A Perl script (
nlfit) implements the multi-resolution fitting
strategy that has been used to map more than 600 brains into stereotaxic
space at the Montreal Neurological Institute. At the heart of this
procedure is a new version of
minctracc, the program that
automatically finds the best non-linear transformation to map one
volumetric data set (stored in MINC format) on to another. The program
uses optimization over a user selectable number of parameters to identify
the best transformation mapping voxel values of the first data set into the