Elastic image registration using improved robust point matching

We present an improved robust point matching (RPM) framework to register elastic deformed images. The framework accelerates its convergence speed and improves its robustness to match landmark sets. In RPM, the initial value of temperature parameter T plays an importance role to estimates the correspondence matrix by deterministic annealing. In this paper, estimation of T based on the location relation of data point sets automatically is presented, and convergence speed of RPM improved greatly. Moreover, estimation principles of the regularization parameters are introduced also. We compared the convergence speed and the robustness of our technique to standard RPM for synthetic data. Finally, we demonstrate the superior performance of the improved RPM to multimodal medical image elastic registration.