MF2MR2: Multi-frequency fusion for accelerated multi-contrast MRI reconstruction

Multi-contrast magnetic resonance imaging (MRI) provides complementary diagnostic information across different pulse sequences, but some contrasts require substantially longer acquisition time, which limits clinical efficiency. Multi-contrast MRI reconstruction (MCMR) addresses this issue by reconstructing a long-acquisition target contrast from highly under-sampled k-space data with a fully-sampled short-acquisition contrast as guidance. However, existing methods still face two major limitations: (1) ineffective modeling of long-range dependencies within and across contrasts, which hinders global anatomical coherence; and (2) insufficient recovery of fine local details, especially subtle high-frequency structures. To address these challenges, we propose MF2MR2, a novel multi-frequency fusion framework that combines the global modeling capability of the Fourier transform with the directional high-frequency decomposition ability of the wavelet transform. The core MF2 module contains three components: (1) a Fourier Fusion Block (FFB) for amplitude-phase based global feature fusion, (2) a High-frequency Refinement Block (HRB) for directional high-frequency detail restoration, and (3) a High-frequency and Fourier Integration Block (HFIB) for structured integration of global and local information. Extensive experiments on BraTS, IXI, and fastMRI show that MF2MR2 consistently outperforms state-of-the-art methods. In particular, on BraTS under 4 ×  acceleration, MF2MR2 achieves 42.04 dB PSNR, outperforming the second-best method by 1.06 dB, while using only 0.929 M parameters. Statistical analysis further confirms that the improvements are significant (p < .05).