Fau-Net: Fourier Attention Based 3D U-Net For Medical Image Registration
Jiong Wu, Yiyuan Wang, Kuang Gong · 2025
Existing U-Net-based unsupervised deformable registration methods often fall short in accuracy due to their limited capacity to capture global features, which are essential for precise registration. To tackle this issue, in this work, we proposed and validated a novel Fourier attention-based 3D U-Net (FAU-Net) for fast medical image registration. The low-level features of the U-Net were first transformed into the frequency domain by leveraging the discrete Fourier transformation (DFT). Afterward, an attention module was introduced to capture the global features from the frequency domain. By conducting the inverse DFT (iDFT), feature maps were recovered from the Fourier spectrum and combined with high-level spatial features from the U-Net for the deformation-field prediction. Evaluations based on two brain magnetic resonance imaging (MRI) datasets show that the proposed framework achieved better registration accuracy and topology preservation than several state-of-the-art unsupervised deformable registration methods.