DC-Reg: A triple-task collaborative framework for few-shot biomedical image registration

Abstract

Deep learning (DL)-based deformable biomedical image registration (DIR) enables automatic information fusion and facilitates rapid diagnosis by aligning multi-source data into a unified coordinate system. However, achieving accurate similarity measurement and obtaining adequate training data pose significant challenges in biomedical image processing tasks. In this paper, we propose a few-shot DIR method that leverages spatial encoding within a triple-task collaborative framework to solve these issues. Firstly, we propose a registration network based on the spatial encoding, which represent images using voxel spatial positions, highlighting the importance of structural information in registration network while reducing modality difference between different images. Secondly, we propose a segmentation network based on data augmentation, which is achieved through the registration network. Specifically, we have designed a contrastive learning based discrimination network to suppress the unreliable augmented training data, which is also our third important component of the collaborative framework. Furthermore, the discrimination network also automatically learns similarity measure for the registration network. By iteratively refining the segmentation, registration, and discrimination networks, we are able to obtain a highly accurate registration model. Our experimental results on four mono-modal and multi-modal datasets demonstrate the effectiveness and superiority of the proposed method.