Meta-Learning–Driven Score-Based Generative Model for 0.5T-to-1.5T MRI Reconstruction

Magnetic resonance imaging (MRI) acquired at low magnetic field strengths typically suffers from reduced signal-to-noise ratios (SNR), which leads to noticeable signal degradation compared with high-field MRI. As a result, reconstructing high-field-like images from low-field MRI data is a challenging task due to the inherently ill-posed nature of the problem. In addition, obtaining paired low-field and high-field MR images is often difficult in practical scenarios.To address these challenges, we propose a novel meta-learning framework with a two-stage mechanism. In the first stage, an optimal-transport-based meta-learner models the degradation process from high-field to low-field MRI and generates pseudo-paired datasets consisting of high-field and low-field images. In the second stage, a base learner solves the inverse problem of recovering high-field-like images from low-field MRI through an iterative regularization strategy, where the learned joint distribution of the pseudo-paired data serves as a prior.Experimental results demonstrate the capability of the proposed approach to generate 1.5T-like images from 0.5T MRI data. Both qualitative visualization and quantitative evaluations, conducted by comparing the reconstructed images with registered real 1.5T images, show that the proposed method produces images with SNR and contrast comparable to those of true 1.5T scans, even under a three-fold acceleration setting. Furthermore, the proposed method achieves superior performance compared with several mainstream approaches, including CycleGAN and Score-MRI.