Low-field MRI reconstruction with hourglass diffusion model and posterior sampling strategy

Motivation: The imaging quality of low-field MRI remains constrained by low signal-to-noise ratio (SNR). Techniques designed to accelerate MR imaging acquisitions may be inapplicable in low-field systems due to low SNR. Goal(s): Our goal is to to jointly reconstruct undersampled low-field k-space data and generate images with reduced noise. Approach: We propose a deep learning model that jointly reconstructs and denoises undersampled low-field images using hourglass transformer and diffusion posterior sampling (LF-DPS) strategy. Results: Our LF-DPS method enhances the quality of low-field images while reducing the acquisition time, and can improve the image quality and diagnostic utility in low-field MRI systems. Impact: An efficiency deep learning based method to accelerate the acquisition in low-field MRI and improve the image SNR