ABSTRACT Accurate medical image registration is crucial for various tasks in medical image analysis. In recent years, Transformer‐based deep learning methods have significantly improved unsupervised registration performance through self‐attention mechanisms and multiscale modeling; however, existing approaches often incur high computational complexity and memory overhead when modeling global and fine‐grained information, making it difficult to balance registration accuracy and efficiency. To address this issue, we propose WHLP‐Net, a pyramid registration network based on discrete wavelets and high‐low frequency Synergistic attention. In the encoding stage, the network adopts the Wavelet‐Convolution‐Fusion (WaveC‐F) module, which uses discrete wavelet transforms to decompose features in the frequency domain, explicitly separating low‐frequency structural information from high‐frequency detail features and fusing them through convolution to form multiscale representations with global structural awareness. In the decoding stage, a High‐Low Synergistic (HLS) attention module is designed to split features along the channel dimension, applying local window‐based attention to the high‐frequency subspace to model local spatial dependencies while leveraging lightweight global average pooling in the low‐frequency subspace to capture global structural information, thereby achieving efficient synergistic modeling of both local and global spatial dependencies. A coarse‐to‐fine pyramid decoding strategy, combined with multilevel deformation field regression and refinement branches, is employed to progressively optimize registration results, particularly for large‐deformation scenarios. We conducted extensive experiments on multiple public datasets. The experimental results show that on the multimodal Abdomen MR‐CT dataset, WHLP‐Net achieves a Dice score of 71.72%, outperforming the classical VoxelMorph and TransMorph methods by 5.92% and 3%, respectively. On the unimodal LPBA and OASIS brain MRI datasets, it achieves Dice scores of 73.4% and 82.72%, respectively. Furthermore, compared with traditional Transformer‐based methods, WHLP‐Net reduces computational complexity by more than 70%, significantly improving computational efficiency while maintaining high registration accuracy.
Fan et al. (Sun,) studied this question.