SBPR-Net: Synchronous Bidirectional Pyramid Registration Network for Medical Image Registration

Jialin Zheng, Wenming Cao, Deliang Lian, Zhiyue Yan

Keywords: deformable image registration, bidirectional registration, deep convolutional neural network, unsupervised learning.

TL;DR: This paper presents SBPR-Net, a novel unsupervised framework for deformable image registration that implements a bridging registration strategy.

Abstract: In recent years, significant advancements have been achieved in deformable image registration. Unlike conventional approaches that directly estimate a flow field at the network output, this paper introduces a novel Synchronous Bidirectional Pyramid Registration Network (SBPR-Net) for unsupervised non-rigid registration. The proposed network implements a bridging registration strategy by simultaneously processing the moving and fixed images, generating two intermediate flow fields that map each image to a common virtual image. The final moving-to-fixed flow field is then derived indirectly through a parameter-free mathematical relationship between these intermediate flow fields. This design enables flexible and efficient deformation estimation of registration. To improve robustness, a novel constraint is introduced to handle virtual image uncertainty, while a pyramid architecture facilitates coarse-to-fine flow field prediction. Extensive experiments on OASIS and IXI brain MRI datasets demonstrate that SBPR-Net outperforms state-of-the-art methods, confirming the effectiveness of the bridging registration strategy.

Submission Number: 41