A consistent deep registration network with group data modeling

Abstract

Medical image registration is a critical process for automated image computing, and ideally, the deformation field from one image to another should be smooth and inverse-consistent in order to bidirectionally align anatomical structures and to preserve their topology. Consistent registration can reduce bias caused by the order of input images, increase robustness, and improve reliability of subsequent quantitative analysis. Rigorous differential geometry constraints have been used in traditional methods to enforce the topological consistency but require comprehensive optimization and are time consuming. Recent studies show that deep learning-based registration methods can achieve comparable accuracy and are much faster than traditional registration. However, the estimated deformation fields do not necessarily possess inverse consistency when the order of two input images is swapped. To tackle this problem, we propose a new deep registration algorithm by employing the inverse consistency training strategy, so the forward and backward deformations of a pair of images can consistently align anatomical structures. In addition, since fine-tuned deformations among the training images reflect variability of shapes and appearances in a high-dimensional space, we formulate a group prior data modeling framework so that such statistics can be used to improve accuracy and consistency for registering new input image pairs. Specifically, we implement the wavelet principle component analysis (w-PCA) model of deformation fields and incorporate such prior constraints into the inverse-consistent deep registration network. We refer the proposed algorithm as consistent deep registration with group data modeling. Experiments on 3D brain magnetic resonance (MR) images showed that the unsupervised consistent deep registration and data modeling strategy yield consistent deformations after switching the input images and tolerated image variations well.