Synthesising 3T DWI from ultra-low-field (64mT) acquisitions using generative diffusion models

Motivation: Diffusion-weighted Imaging (DWI) at very-low fields like the 0.064 Tesla Hyperfine Swoop is limited by low signal-to-noise ratio (SNR), impeding clinical application. Goal(s): This study aims to enhance DWI at such low fields by creating synthetic high-field images using pre-trained neural networks. Approach: The Diffusion Probabilistic Model (DPM), an advanced generative AI, will be trained on high-quality 3T DWI images to learn their distribution. Low-field DWI images guide the DPM to conditionally synthesize high-quality images. Results: With a well-trained DPM, we aim to produce high-quality, synthetic 3T-like DWI images that mirror the original low-field ones, bypassing the need for paired training data. Impact: The method enhances DWI image quality at very-low field strength in an unsupervised manner, eliminating the need for paired high-field and low-field data, thus expanding training data availability. Zero-shot image reconstruction enhances its generalizability for diverse tasks.