An Unsupervised Non-rigid Registration Network for Fast Medical Shape Alignment.

Abstract

Medical shapes alignment can provide doctors with abundant structure information of the organs. As for a pair of the given related medical shapes, the traditional registration methods often depend on geometric transformations required for iterative search to align two shapes. To achieve the accurate and fast alignment of 3D medical shapes, we propose an unsupervised and nonrigid registration network. Different from the existing iterative registration methods, our method estimates the point drift for shape alignment directly by learning the displacement field function, which can omit additional iterative optimization process. In addition, the nonrigid registration network can also adapt to the geometric shape transformations of different complexity. The experiments on two types of 3D medical shapes (liver and heart) at different-level deformations verify the impressive performance of our unsupervised and nonrigid registration network.Clinical Relevance-This paper achieves the real-time medical shape alignment with high accuracy, which can help doctors to understand the pathological conditions of organs better.