Abstract: Hyperpolarized ¹²⁹Xe and ¹H magnetic resonance imaging (MRI) visualize lung structure and ventilation, enabling early diagnosis. However, ¹²⁹Xe imaging faces low signal noise ratio (SNR), accelerated sampling, and multinuclear heterogeneity. Existing methods have clear limits: regression models outputs are over-smoothed, losing high-frequency details; mainstream generative models are noise-sensitive, artifact-prone, and slow; multinuclear heterogeneity is unmodeled, underusing multimodal complementarity. We propose a flow-matching framework for multinuclear lung MRI reconstruction. It integrates regression and generative strengths for better metric–fidelity trade-off, efficient inference, noise robustness, and improved complementary information utilization, supporting accelerated clinical multinuclear MRI.

Key words: Flow Matching Model, Multi-Nuclear MRI, Image Reconstruction, Deep Learning