Measurement-Consistent Low-Field MRI Enhancement via k-Space Noise Modeling

Abstract

Low-field magnetic resonance imaging (MRI) has gained increasing attention as a cost-effective and portable alternative to high-field MRI, particularly in resource-limited conditions. However, its inherently low magnetic field strength leads to reduced signal intensity and a low signal-to-noise ratio (SNR), resulting in severe image degradation. To address these challenges, this work proposes a framework for low-field MRI enhancement that jointly considers data synthesis, model design, and performance evaluation in a measurement-consistent manner. In particular, we first propose a realistic low-field MRI synthesis strategy that estimates noise statistics from k-space measurements and injects noise into high-field k-space. By separately modeling the real and imaginary components, the proposed pipeline enables more faithful generation of paired low- and high-field MRI data for supervised learning. Second, we establish a unified benchmark that supports reliable performance evaluation, which includes a wide range of representative restoration methods, together with a real-world dataset consisting of low-field MRI scans and corresponding high-field References acquired from the same subjects. Finally, we develop the noise-aware guidance network (NAGNet), which incorporates an adaptive noise awareness (ANA) branch to explicitly predict spatially variant noise and guide the denoising process in a plug-and-play manner. Extensive experiments demonstrate that NAGNet achieves state-of-the-art performance, which exhibits strong generalization across different datasets and various magnetic field strengths. Code and datasets are available at https://github.com/6767lq/NAGNET