Abstract: Medical image registration is essential for several post-processings.Even though the existing single-stream or dual-stream network structures based on convolution and Transformer can achieve the promising results,it is still difficult to make a compromise between the registration performance and computational efficiency.To deal with this issue,this paper proposes an efficient registration network EFormer which mainly consists of the frequency division module(FDM) and broad attention module(BAM).Specifically,stacking several FDMs in encoder and decoder to mimic the roles of dual-branch network for extracting both local and global information can significantly improve the computation efficiency,using BAM to enhance the transmission of local information in multiple FDMs can preserve significant semantic features to promote the registration performance.The qualitative and quantitative comparisons with state-of-the-art methods on three datasets demonstrate that the Dice score,ASSD,HD95 and the ratio of negative Jacobian determinant of the proposed EFormer is improved at least by 1.3%,2.6%,0.6% and 95% respectively.In addition,using EFormer-tiny,the computation efficiency(Flops) is improved by 14%,showing that the proposed EFormer can achieve the best registration results in attention-based networks with the fastest computation speed.

Key words: Medical image registration, Frequency division, Broad attention, Efficient Transformer