A new multiscale registration method for medical image

Mutual information (MI) is a well accepted similarity measure for image registration. However, MI based registration faces the challenges of high computational complexity, low registration efficiency and high likelihood of being trapped into local optima due to an absence of spatial information. In this paper, we propose a new multiscale registration framework based upon an edge preserving total variation L1 norm (TV-L1) scale space representation. Our scale space is constructed by selecting edges and contours of an image according to the geometric size rather than the intensity values of the image features. This ensures more meaningful spatial information for MI based registration. Furthermore, we design an optimal estimation of the TV-L1 parameter in our framework by training and minimizing the transformation offset between the images for automated registration. We validated our method on both simulated mono- and multi-modal medical datasets with ground truth and temporal clinical studies from a combined PET/CT scanner.