A probabilistic denoising diffusion-based framework for even higher accelerated quantitative MRI

Motivation: Fast quantitative MRI using highly accelerated acquisitions like in MRF comes at the cost of severe aliasing artifacts that needs to be resolved Goal(s): Addressing undersamling artifacts and quantifying uncertainties in quantitative maps to pave the way to even shorter acquisitions e.g. in MRF Approach: Introducing the first probabilistic diffusion-based framework for the example of MRF reconstruction, advancing state-of-the-art-deep learning techniques for more accurate quantitative mapping with tools to assess uncertainties Results: Quantitative and qualitative evaluations show that our diffusion-based approach outperforms state-of-the-art in producing more accurate tissue parameters. Uncertainty maps exhibit correlations between areas of large variance with areas of large errors. Impact: Our proposed approach enables the efficient use of Improved Denoising Diffusion Probabilistic Models for reconstructing highly accelerated quantitative MRI acquisitions, such as Magnetic Resonance Fingerprinting, leading to more accurate tissue parameter estimations.