Self‐supervised multi‐contrast MRI reconstruction via mask‐guided frequency‐domain fusion

Background Multi‐contrast magnetic resonance imaging (MC‐MRI) enables detailed tissue characterization by acquiring multiple contrast‐weighted images from the same anatomical region. Redundancy across contrasts can improve reconstruction from under‐sampled data. However, conventional deep learning approaches rely on fully‐sampled, contrast‐aligned datasets, which require long acquisition times. Extended scanning durations not only increase patient burden but may also introduce physiological risks, limiting the practical adoption of accelerated MC‐MRI in clinical settings. Purpose This study introduces Self‐MF, a self‐supervised MC‐MRI reconstruction framework that leverages contrast redundancy without requiring fully sampled target‐contrast labels. The framework aims to provide a high‐fidelity, scalable solution for reconstructing under‐sampled multi‐contrast images while preserving anatomical details and contrast fidelity. Methods Self‐MF implements a contrast‐aware self‐supervised learning strategy that randomly partitions the acquired target k‐space into input and held‐out subsets, allowing the model to learn joint anatomical priors from under‐sampled data without fully sampled target‐contrast labels. A mask‐guided frequency‐domain conditioning module constructs reference‐conditioned auxiliary frequency channels from auxiliary contrasts and acquired target measurements, enhancing the recovery of fine structural features during score‐based reconstruction. Experiments were conducted on public datasets (IXI, fastMRI) and a clinical multi‐coil dataset (SIAT), spanning multiple acceleration factors and undersampling patterns to assess robustness and generalizability. Results Quantitative and qualitative evaluations demonstrate that Self‐MF outperforms existing representative reconstruction methods across different contrasts, datasets, and acceleration factors. On the SIAT multi‐contrast dataset, Self‐MF achieved PSNR/SSIM values of 35.16/0.8870, 33.46/0.9054, and 35.31/0.9187 for T1WI, T2WI, and PDWI reconstruction under 2D random undersampling at R = 4, respectively. On the fastMRI brain dataset, Self‐MF achieved 41.15 dB/0.9637 and 35.02 dB/0.8814 for T2WI reconstruction under R = 4 and R = 8, respectively, consistently outperforming the compared classical, learning‐based, and diffusion‐based baselines. Conclusions Self‐MF provides an effective, self‐supervised solution for high‐fidelity MC‐MRI reconstruction without fully sampled target‐contrast labels. By combining contrast‐aware self‐supervision, frequency‐domain conditioning, and diffusion‐based reconstruction, the framework achieves superior image quality and robust anatomical preservation, representing a significant advancement for multi‐contrast MRI in both research and clinical practice.