DDMamba: A Dual-Domain Mamba for Multi-Modal Magnetic Resonance Imaging Reconstruction With Fourier Fusion

Although reconstruction using shared information between multi-modal magnetic resonance (MR) images has shown promising results, there remains a need to develop new methods for achieving more efficient and reliable MRI reconstruction.

A Dual-domain Mamba Network (DDMamba) for multi-modal MRI reconstruction with Fourier fusion is proposed to solve the problem of reconstructing MR images that have been significantly under-sampled. DDMamba includes three key modules: Spatial Mamba, Wavelet Mamba, and Fourier fusion. Spatial Mamba refines the spatial information of input features through a multi-scale feature extraction block, providing richer spatial features for the state space model. Wavelet Mamba introduces a novel scanning strategy, namely, Wavelet 2D-Selective-Scan, which establishes dependency relationships from low to high frequencies. The Fourier domain operation uses global characteristics to enhance spatial and wavelet information fusion.

Extensive experiments conducted on public data sets, NAMIC, BraTS, and fastMRI, have demonstrated the superior performance of the proposed method in multi-modal MRI reconstruction. Notable improvements are observed in both quantitative metrics and visualizations of the proposed method when compared to state-of-the-art approaches.

The proposed Dual-domain Mamba network integrates spatial and wavelet domain features through Fourier fusion, effectively optimizing multi-feature representation. This method significantly enhances the reconstruction of multi-modal single-coil MRI, addressing the demand for efficient and reliable imaging solutions in clinical practice.