Denoising Diffusion Probabilistic Models for Magnetic Resonance Fingerprinting

Magnetic Resonance Fingerprinting (MRF) is a time-efficient approach to quantitative MRI, enabling the mapping of multiple tissue properties from a single, accelerated scan. However, achieving accurate reconstructions remains challenging, particularly in highly accelerated and undersampled acquisitions, which are crucial for reducing scan times. While deep learning techniques have advanced image reconstruction, the recent introduction of diffusion models offers new possibilities for imaging tasks and their application in the medical field. Notably, diffusion models have only recently been explored for quantitative MRI and remain largely unstudied for MRF reconstruction. In this work, we propose a conditional diffusion model to reconstruct MRF data, demonstrating its potential for accurate quantitative MRI from accelerated acquisitions. Our findings are supported by qualitative and quantitative comparisons on in-vivo brain scan data, demonstrating that the proposed approach can outperform established deep learning and classical compressed sensing algorithms for MRF reconstruction in highly accelerated regimes. In our experiments, our approach achieves reductions in mean percentage errors of at least 0.71\% and 2.15\% for T1 and T2 reconstructions, respectively, at an acceleration factor of R=5. A range of ablation studies also explore strategies to improve computational efficiency of our approach.