A novel multi-layer framework for non-rigid image registration

Brain magnetic resonance (MR) images consist of different structures and features when they are observed at different scales and layers. Conventional non-rigid brain MR image registration methods mainly estimate the optimum transformation by relying on the information of a single layer and this can lead to the loss of information contained in other layers. In this paper, we propose a multi-layer framework for non-rigid brain MR image registration with different kinds of features extracted from different layers. The input images are factorized into three layers: global intensity layer, texture information layer and local anatomical layer. The generalized survival exponential entropy based mutual information (GSEE-MI), multi-scale brainton features and rotation invariant feature transform (RIFT) are used to represent the global intensity layer, texture information layer and local anatomical layer respectively. Information extracted from all layers is then embedded into a new similarity measure function. The role of each layer is identified through systematic experiments and it is shown that information conveyed by different layers is complement with each other. The proposed framework exhibits significant improvement of registration accuracy as compared with other widely used registration methods on the real 3D databases obtained from IBSR.