The processing stream for spatial normalization is presented schematically in Fig. 2. The general SN algorithm is applicable to both global and regional transformation methods by altering feature matching and transform strategies. Global spatial normalization is the simplest to describe since it has the fewest features to match. For Talairach spatial normalization the target-brain features are derived from the 1988 Talairach atlas. In most cases processing is iterative with a goal of minimizing differences between features in transformed and target brains. This is achieved by adjusting transform parameters between processing steps, following analysis of residual feature differences. For automated methods, the feature difference is often converted to a mean square error (MSE) and an MSE minimization method used to drive the transformation [5,26,57]. For manual methods feature comparison is done by visual inspection, and the transform is adjusted interactively to minimize feature differences (see example). In high degree-of-freedom regional deformation methods, both feature matching and transformation are performed using a global-to-regional scheme. This hierarchical multiresolution approach deals with large features first, proceeding to smaller features, and down to the limiting resolution of the processing algorithm if desired [5,6,26,31,34].
Was this article helpful?