Topology Preserving Image Transformation through Mean Shift Iteration

Transformations based on radial basis functions expansion are not invertible and topology is not preserved in elastic image registration with large deformation.Under the precondition of correct correspondences of control points,a novel topology-preserving transformation is presented in this paper which warps images for large deformation using radial basis functions and Mean Shift iterations.At first,additional points were tacked at topology non-preserving regions where Jacobian values are negative.Mean Shift iteration was adopted to modify additional target points.The optimal additional point pairs,which resulted in positive Jacobian values of transformation function and improved registration accuracy,were selected.They were combined with original control point set to construct topology preserving transformation functions with radial basis functions.The image registration accuracy was improved using the extended control point set.We compared our method with traditional ones and demonstrated that the proposed method was advantageous on preserving topology of deformation field.Experiments of artificial images and real medical images showed the feasibility of our method.