Mutual information based non-rigidmouse registration using a scale-space approach

We propose a scale-space based approach to non-rigid small animal image registration. Scale-space theory is based on generating a family of images by blurring an image with Gaussian kernels of increasing width. This approach can be used to extract features at varying levels of detail from an image. We define the scale-space feature vector at each voxel of an image as a vector of intensities of the scale- space images at that voxel. We generate scale-space images of the target and template images, and extract their corresponding scale- space feature vectors at each voxel. The extracted feature vectors are aligned using mutual information based non-rigid registration to simultaneously align global structure as well as detail in the images. We represent the displacement field in terms of the discrete cosine transform (DCT) basis, and use the Laplacian of the displacement field as a regularizing term. The DCT representation of the displacement field simplifies the Laplacian regularization term to a diagonal, thus reducing computational cost. We apply the scale-space registration algorithm on mouse images obtained from two time points of a longitudinal study, and compare its performance with that of a hierarchical multi-scale approach. The results indicate that scale- space based registration gives better skeletal as well as soft tissue alignment compared to the hierarchical multi-scale approach.