Memory-Efficient Image Reconstruction using Diffusion Models for Accelerated 3D Non-Cartesian UTE imaging

Motivation: 3D radial sampling is crucial in UTE sequences for effectively capturing short-T2 tissue such as bone. Non-Cartesian reconstructions pose difficulties due to high memory demands and computational times. Goal(s): This work aims to develop a diffusion-model to achieve high-quality images from undersampled k-space data within reasonable reconstruction time. Approach: Our approach integrates a memory-efficient neural-network, employing Heun's efficient sampling and conjugate gradient-based data consistency. Results: The proposed reconstruction yields high SSIM and PSNR values with good generalizability across acceleration factors and body regions, demonstrating its effectiveness for 3D non-Cartesian reconstruction allowing shorter scan times. Impact: We propose a memory-efficient diffusion model for reconstructing accelerated 3D radial UTE acquisitions, enabling high-quality, and reliable reconstructions while reducing the reconstruction time of common deep-learning methods. The model generalizes well across body parts, supporting various applications and acceleration factors.