Extremely Accelerated MRI by way of Implicit Neural Illustration Guided Posterior Sampling of Diffusion Fashions
Authors: Jiayue Chu, Chenhe Du, Xiyue Lin, Yuyao Zhang, Hongjiang Wei
Summary: Reconstructing high-fidelity magnetic resonance (MR) pictures from under-sampled k-space is a generally used technique to cut back scan time. The posterior sampling of diffusion fashions primarily based on the true measurement information holds vital promise of improved reconstruction accuracy. Nonetheless, conventional posterior sampling strategies usually lack efficient information consistency steering, resulting in inaccurate and unstable reconstructions. Implicit neural illustration (INR) has emerged as a strong paradigm for fixing inverse issues by modeling a sign’s attributes as a steady operate of spatial coordinates. On this research, we current a novel posterior sampler for diffusion fashions utilizing INR, named DiffINR. The INR-based element incorporates each the diffusion prior distribution and the MRI bodily mannequin to make sure excessive information constancy. DiffINR demonstrates superior efficiency on experimental datasets with exceptional accuracy, even underneath excessive acceleration components (as much as R=12 in single-channel reconstruction). Notably, our proposed framework could be a generalizable framework to unravel inverse issues in different medical imaging duties.
