HFGN: High-Frequency residual Feature Guided Network for fast MRI reconstruction

Magnetic Resonance Imaging (MRI) is a valuable medical imaging technology, while it suffers from a long acquisition time. Various methods have been proposed to reconstruct sharp images from undersampled k-space data to reduce imaging time. However, these methods hardly reconstruct high-quality aliasing-free Magnetic Resonance (MR) images with clear structures, especially in high-frequency components. To address this problem, we propose a High-Frequency residual feature Guided Network (HFGN) for fast MRI reconstruction. HFGN uses a sub-network, High-Frequency Extraction Network (HFEN), to learn the difference between the U-Net reconstruction result and the ground truth, then uses the learned features to guide the reconstruction of the network. In the reconstruction network, we propose Residual Channel and Spatial Attention block (RCSA), which uses frequency domain and image domain convolution branching to learn the global and local features of the image simultaneously. The experiment results under different acceleration rates on different datasets demonstrate that our proposed method surpasses the existing state-of-the-art methods.