Frequency Filtering Network for Undersampled MRI Reconstruction

Magnetic resonance imaging (MRI) is an imaging technology widely used in medical clinical diagnosis. Nevertheless, it always takes a long imaging time to obtain the high spatial resolution MR image. Image reconstruction has been playing a crucial role for accelerated MRI. Recently, deep neural networks show potential to significantly speed up MRI. However, these deep learning methods usually learn the feature information in the spatial domain. Although some works have attempted to utilize the frequency information from the k-space to reconstruct MR images, they still suffer from the inability to consider the global information. To address these limitations, in this work, we propose a new residual filter channel attention network (RFCAN) for undersampled MRI reconstruction. Specifically, the proposed model consists of a novel filter channel attention (FCA) in the frequency domain and the multiscale cascade convolution in the spatial domain to learn both the dual-domain features. This model employs the filtering instead of the traditional self-attention mechanism in Transformer to learn k-space information. It can avoid the limitations of the Transformer window to better learn the global information. We also introduce a frequency channel attention to fine-tune the filtering results to better balance the global and local information. Experiment results show that our model consistently outperforms other state-of-the-art methods and can achieve high-quality MRI reconstruction.