Robust 3D deformation field estimation by template propagation

Abstract

A new robust method to automatically determine a 3D motion vector field for medical images in the presence of large deformations is proposed. The central idea of this approach is template propagation. Starting from an image position where valid starting estimates are known, small sub-volumes (templates) are registered rigidly. Parameters of successfully registered templates serve as starting estimates for its neighbors. The registration proceeds layer by layer until the relevant image volume is covered. Based on this principle, a template-based registration algorithm has been implemented. Using the resulting set of corresponding points, the parameters of a non-rigid transformation scheme are determined. The complete procedure has been validated using four MR image pairs containing considerable deformations. In order to obtain an estimate for the accuracy, homologous points determined by template propagation are compared to corresponding landmarks defined by an expert. For landmarks with sufficient structure, the average deviation is well below the voxel size of the images. Because of the larger number of homologous points available, transformations incorporating the output of template propagation yielded a larger similarity between the reference image and the transformed image than an elastic transformation based on landmark pairs.