Medical image registration based on generalized N-dimensional principal component analysis (GND-PCA) and statistical shape deformation model

Abstract

This paper presents a fast and accurate image registration method for high dimensional images. The method uses a statistical shape deformation model to represent deformation fields which warp an individual image to a selected template image. The statistical shape deformation model is built by the generalized N-dimensional principal component analysis (GND-PCA) with training samples of deformation fields, which deform the individual sample images to the selected template image. The statistical deformation model can be built with fewer samples and can represent individual deformation fields effectively by a small number of parameters, which is used to rapidly estimate the deformation field between the template image and a new individual image. The estimated deformation field is used to warp the individual image, and the warped image is close to the template image. The shape difference between the warped individual image and the template is estimated by an image registration algorithm, e.g., HAMMER. The proposed method has been validated by 3D MR brain images.