Diffusion Probabilistic Generative Models for Accelerated in-NICU, Permanent Magnet Neonatal MRI Reconstruction

Motivation: Accelerating lower-field MRI in the neonatal-intensive-care-unit may reduce motion artifacts and increase accessibility to a wider range of patients. However, parallel imaging and machine learning reconstruction models trained on adult MRI do not apply in the neonatal setting. Goal(s): This work accelerates single-channel neonatal MRI with diffusion-probabilistic-generative models trained from limited and noisy data collected on a permanent magnet system. Approach: The proposed training method combines datasets from multiple contrasts and orientations with class embeddings and applies a self-supervised denoiser before training. Diffusion posterior sampling reconstructs images from under-sampled k-space. Results: Our method enables 1.5x reduction in scan-time using a single-channel. Impact: The improvement in acquisition speed of T1 and T2 weighted lower field neonatal MRI protocols using diffusion-probabilistic-generative models, trained with methods designed to handle the noisy, limited data, improves accessibility of MRI to patients in the neonatal-intensive-care-unit.