Cross-Field MRI Synthesis using Flow Matching in Multiple Sclerosis Imaging

Portable ultra-low field (pULF, 64mT) MRI presents a transformative opportunity for accessible neuroimaging, particularly in underserved or point-of-care settings. However, its limited resolution and signal-to-noise ratio pose challenges for clinical applications such as multiple sclerosis (MS), where reliable visualization of anatomical structure and lesions is essential to gauge disease progression and therapeutic efficacy. In this study, we propose SynthXFM (Synthesis via Cross-field Flow Matching), a novel flow-based generative model for cross-field MRI synthesis between ultra-low-field (64 mT) and high-field (3T) MRI. SynthXFM leverages a deterministic flow matching framework to learn continuous, structure-preserving mappings from noise to image space, conditioned on pULF scans. Compared to conventional generative models such as Pix2Pix GAN, SynthXFM produces images with sharper anatomical detail, realistic tissue textures, and significantly better control over hallucination artifacts. Quantitative evaluations on a dataset of 102 training and 20 test subjects show that SynthXFM achieves superior performance in PSNR, MS-SSIM, and LPIPS metrics. Segmentation-based evaluation further confirms improved anatomical fidelity, with higher Dice scores across multiple brain regions when compared to high-field reference MRI. Additionally, we demonstrate the ability to perform high-field to ultra-low-field synthesis using the same approach. These results highlight the potential of SynthXFM to close the quality gap in pULF MRI and support clinically meaningful neuroimaging in low-resource environments.