Non-linear 2D and 3D Registration Using Block-Matching and B-Splines

We developed a non-linear registration technique to align images that feature anatomical variabilities. The algorithm is based on a block-matching technique that identifies a sparse displacement vector field from the iconic features of two images. Subsequently, the displacement vectors are used as sampling points to estimate a parametric non-linear transformation that is represented by a tensor product of B-Splines. The B-Spline transformation estimation approximates the correspondences while minimizing the second order derivatives in the transformation function. The block-matching and the transformation estimation are then iterated in a multiscale framework to improve robustness and accuracy. Experiments on 2D histological slices and 3D MR images show qualitatively good results.