High-resolution ultra-low-field MRI with SNRAware denoising

Abstract

Ultra-low-field (ULF, <0.1 T) magnetic resonance imaging (MRI) systems offer advantages in cost, portability, and accessibility, but their current utility is still limited by low signal-to-noise ratio (SNR). Deep learning (DL)-based denoising has emerged as a potential strategy to mitigate this limitation. In this work, we present a systematic evaluation of a high-performance DL denoising model trained using the SNRAware framework and applied to 88 mT and 72 mT data. Using a series of controlled experiments, we assessed model performance as a function of spatial resolution, coil impedance matching, readout bandwidth, input noise level, k-space undersampling, anatomy, image contrast, and scanner platform, and compared against analytical denoising algorithms. The model consistently increased the effective SNR of ULF acquisitions, enabling images with nominal spatial resolutions comparable to those commonly used in clinical 3 T protocols. Residual analyses indicated that the model predominantly removed stochastic noise while preserving underlying signal structure. At the same time, the results highlight some constraints: denoising performance remains dependent on the starting SNR of the acquisition, and training-domain mismatch influences behavior under certain artifact conditions. These findings suggest that DL-based denoising can significantly expand the practical capabilities of ULF MRI, while emphasizing potential benefits from hardware-software co-optimization and the need for rigorous clinical validation to determine the diagnostic value of denoised images.

Figures (9)

• Figure 1: SNRAware training. Complex white noise is first sampled. Noise augmentation takes in the g-factor map to generate realistic, spatially variant, correlated noise. This is added to high-SNR data to create low-SNR images. The paired high-/low-SNR data is used to compute loss and optimized model parameters. In inference, complex, low-SNR images and g-factor maps are inputted into the trained model, which produces denoised complex images as final outputs. Further details can be found in Refs. Xue2025Xue2025b.
• Figure 2: Portable Halbach scanners employed in this work. (a) 88 mT elliptical system for head and extremity imaging (NextMRI). (b) 72 mT circular system for extremity imaging (Physio I).
• Figure 3: Representative results from Experiment 1. Top: axial knee magnitude images reconstructed using FFT and SNRAware at 3 T clinical resolution (0.3 $\times$ 0.3 $\times$ 3 mm$^3$). Middle: residual magnitude images relative to the raw FFT reconstruction. Bottom: magnitude of the central $k$-space slice for each reconstruction. The SNR (noise standard deviation) values displayed in the reconstructions were computed within the ROI indicated by the white (yellow) boxes.
• Figure 4: Representative results from Experiment 2. Sagittal knee images at isotropic resolutions of 2 mm, 1.5 mm, and 1 mm, and ACR phantom images at 0.2 $\times$ 0.2 mm$^2$ in-plane resolution. Columns show raw FFT magnitude reconstructions, SNRAware-denoised images, and residual magnitude images relative to the raw FFT reconstruction. The SNR (noise standard deviation) values displayed in the reconstructions were computed within the ROI indicated by the white (yellow) boxes.
• Figure 5: Representative results from Experiments 3–6. ACR phantom reconstructions comparing raw FFT and SNRAware under varying acquisition conditions. (a) Good (–20 dB) and poor (–3 dB) impedance matching with 2 ms and 12 ms readouts. (b) Baseline (1.2$\times$) and elevated (6.2$\times$) noise levels. (c) Fully sampled and 60 % Partial Fourier acquisitions. Zoomed-in panels display the high-resolution region of the ACR phantom.