Fast unconditional diffusion model for accelerated MRI reconstruction

Accelerated magnetic resonance imaging (MRI) reconstruction from undersampled k -space data is a challenging inverse problem that has attracted significant attention in the MRI community. Diffusion models have recently emerged as a promising solution for MRI reconstruction, as they can generate high-quality samples while maintaining sample diversity. However, the inference process of diffusion models is computationally expensive, requiring thousands of steps to ensure the quality of the generated samples, which can take tens of minutes to complete. To address this issue, we propose a novel fast diffusion model for MRI reconstruction, termed FDMR, which aims to accelerate the inference process and improve reconstruction quality. The FDMR framework consists of two main components: the adversarial training of the denoising diffusion GAN and the three-stage inference framework. The adversarial training process is used to train the denoising diffusion GAN with large steps, learning an unconditional diffusion prior and embedding a deep generative prior. The proposed three-stage inference framework includes fast diffusion generation, early stopped deep generative prior adaptation, and diffusion refinement, aiming to accelerate the inference process and improve the reconstruction quality. Extensive experiments demonstrate that FDMR can achieve superior reconstruction accuracy compared to state-of-the-art diffusion methods, yet it operates 4-10 times faster, enabling the reconstruction within just 8 s.