Deformable medical image registration based on wavelet transform and linear attention

Medical image registration, which is the process of aligning two medical images to ensure spatial consistency, is extensively utilized in surgical navigation, lesion localization, and other fields. In recent years, Transformer-based models have achieved state-of-the-art image registration results. However, these models often lose significant information during image downsampling, and the window division operation in the window attention module disrupts the integrity of the anatomical structure, which limits the model's ability to capture global features. To address these problems, this study introduces an image registration model based on the wavelet transform and linear attention. Specifically, the discrete wavelet transform is leveraged to decompose the image, to provide more features to the model and harness the rich high-frequency information achieved from the wavelet transform, to effectively mitigate the information loss arising from downsampling. Wavelet transform and linear attention are combined to propose a linear wavelet self-attention (LWSA) module. Compared with other attention modules, this unique linear attention calculation method in LWSA reduces the computational complexity and significantly expands the receptive field of the module, ensuring the integrity of the anatomical structure. Additionally, the introduction of the wavelet transform allows the model to pay differential attention to different areas, thereby improving the model's performance. The efficacy of the proposed model in registration tasks was validated through comparative experiments with other models on three publicly available brain imaging datasets.