Spatial deformation models for non-rigid image registration

Spatial deformation models are used to regularize image registration such that they prevent physically and anatomically unlikely transformations. It is often assumed that optimal models are obtained by modeling deformation properties of real tissues. However, this is not exactly true, because external forces, which drive the registration, in general differ from forces which in reality deformed the anatomy. In order to develop better spatial deformation models, it is necessary to consider these differences. In this work we focus on convolution based models. We analyze advantages and disadvantages of two most commonly used spatial deformation models, i.e. elastic model, and incremental model, and two widely used convolution kernels: an elastic kernel and a Gaussian kernel. The result of this work is a new combined elastic-incremental model, suitable for non-rigid registration of medical images.