Diffusion Probabilistic Models for Image Formation in MRI

Diffusion probabilistic modeling has recently emerged as a state-of-the-art framework in MRI image-formation tasks. Two mainstream tasks in this domain are image reconstruction from undersampled k-space acquisitions with the purpose of accelerating MRI exams, and image translation to impute missing sequences for broadening the scope of multi-contrast MRI protocols. Diffusion models, known for their exquisite capability to generate high-fidelity images, have demonstrated great promise in solving the ill-posed inverse problems associated with these tasks. In the context of reconstruction, diffusion models have shown prowess in recovering high-quality MR images from heavily undersampled acquisitions, to enable significant reductions in scan times. In the context of translation, they have shown superior quality in imputed images of missing sequences, to ensure availability of comprehensive multi-contrast MRI protocols without the need for additional exams per patient. This chapter provides a comprehensive overview of the theoretical foundations, practical implementations, and recent advancements in the use of diffusion models for these pivotal MRI tasks, highlighting the potential of this deep learning framework to transform clinical imaging practices. Through detailed discussions and illustrative examples, we explore how diffusion models can bridge existing gaps in MRI technology, paving the way for faster, more accurate, and comprehensive imaging solutions.