Registration: 2D/3D rigid registration

We introduce how to leverage deep learning models to solve problems in medical image registration. Since medical image registration is a broad topic, we focus on the two popular directions in addressing medical image registration with deep learning, namely deep iterative registration and transformation estimation based registration. For each direction, we present two categories of approaches and introduce the choices of object functions, as well as example models for particular registration applications. In the case study, we limit the problem to the 2D/3D rigid registration and investigate how the problem can be addressed in practice. We give a formal introduction to the 2D/3D rigid registration problem setting and propose a Point-Of-INterest Tracking (POINT) based registration approach. Specifically, we provide details on how to design a U-Net based Siamese network for tracking, and on how to establish 3D point correspondences from the 2D tracking maps. We show that the proposed approach achieves better registration performance while being more computationally efficient than reinforcement learning based and optimization-based approaches.