Motivation: Low-field MRI produces less acoustic noise, making it more suitable for newborns. However, low-field MRI, especially diffusion-weighted imaging (DWI), suffers from low SNR and reduced contrast, which limits its clinical effectiveness in diagnosis. Goal(s): To enhance the image quality of DWI at 0.35 T. Approach: A self-training two-stage framework was proposed, which extracted the structural and contrast representations of infant DWI from 0.35T and 1.5T, and aligned the above features using a proposed contrastive loss function. Results: Experiments on clinical infant brains achieved a PSNR of 23.88 and a SSIM of 0.8944, outperforming other networks and demonstrating the superior performance of our method. Impact: The proposed method improved the image quality of DWI at low-field without collecting paired LF-HF data. It may promote accurate diagnosis using low-field MRI for DWI.
Chen et al. (2025) studied this question.