BPR-Net: Bridging Pyramid Registration Network for Unsupervised Medical Image Registration

Recently, many deep learning-based methods have been proposed to tackle deformable image registration and achieved remarkable success. In this paper, comparing with those existing deep learning-based methods which generate the flow filed directly in the end of the network, we propose a novel Bridging Pyramid Registration Network (BPR-Net) for unsupervised non-rigid registration, which enables the same registration task to be carried out simultaneously from two direction and generates two deformation fields which respectively represent the registration from both sides to the middle state. To obtain the desired deformation field, we figure out the relationship between the output fields and our desired field by a mathematical model. To ensure the robustness and efficiency of the strategy, we introduce a new constrain to handle the uncertainty problem of the middle state. After that, we proposed the middle generation module to fuse the contextual features which is able to predict the middle state dynamically. Such strategy provides the network with ability to achieve accurate registration between two volumes in a more flexible and efficient ways. Extensive experiments on three public brain MRI datasets show that the proposed BPR-Net outperforms the existing state-of-the-art deformable image registration methods which shows the feasibility and effectiveness of the bridging registration strategy.