MR imaging in the low-field: Leveraging the power of machine learning

TL;DR

Low-field MRI ($<1\,\mathrm{T}$) and ultra-low-field MRI ($<0.1\,\mathrm{T}$) face SNR and time constraints; the paper surveys ML-driven solutions for reconstruction, denoising, and super-resolution to mitigate these limitations. It covers training paradigms—from supervised to self-supervised (e.g., $${\text{Noise2Noise}}, {\text{Noisier2Noise}}, {\text{SSDU}}$$$)—and physics-informed and unrolled approaches that leverage forward models like $${\mathbf{s}=\mathbf{E}\boldsymbol{\rho}_{\mathrm{true}}+\boldsymbol{\eta}}$$. The discussion highlights concrete methods such as AUTOMAP, SRDenseNet, and joint image/field-map reconstructions (e.g., SH-Net) and discusses domain adaptation for cross-field data. The work argues that ML-enabled low-field MRI could substantially expand access and reduce costs, but stresses the need for robust validation, standardized protocols, and careful handling of artifacts and safety to realize its clinical impact.

Abstract

Recent innovations in Magnetic Resonance Imaging (MRI) hardware and software have reignited interest in low-field ($<1\,\mathrm{T}$) and ultra-low-field MRI ($<0.1\,\mathrm{T}$). These technologies offer advantages such as lower power consumption, reduced specific absorption rate, reduced field-inhomogeneities, and cost-effectiveness, presenting a promising alternative for resource-limited and point-of-care settings. However, low-field MRI faces inherent challenges like reduced signal-to-noise ratio and therefore, potentially lower spatial resolution or longer scan times. This chapter examines the challenges and opportunities of low-field and ultra-low-field MRI, with a focus on the role of machine learning (ML) in overcoming these limitations. We provide an overview of deep neural networks and their application in enhancing low-field and ultra-low-field MRI performance. Specific ML-based solutions, including advanced image reconstruction, denoising, and super-resolution algorithms, are discussed. The chapter concludes by exploring how integrating ML with low-field MRI could expand its clinical applications and improve accessibility, potentially revolutionizing its use in diverse healthcare settings.

Figures (9)

• Figure 1: Comparison of a high-field ($3\,\mathrm{T}$) and a corresponding simulated ultra-low-field (50mT) MR image. Lower field systems typically only achieve a decreased spatial resolution (downsampled), are exposed to strong B0-inhomogeneities (B0 distortion), poor SNR (noise) and use receiver coils with high Q-factors (coil profile). The upper left image is an example taken from the fastMRI brain dataset zbontar2018fastmri, the intermediate images were obtained by applying the just mentioned different degradation steps. The lower right image is taken from de2022deep and was acquired with the $50\,mT$ ultra low-field scanner described in o2021vivo. Image courtesy of Andrew Webb.
• Figure 2: Different training techniques for a simple image-denoising example. Supervised training requires access to input-target image pairs and the network $\mathrm{Net}_{\theta}$ is trained to estimate the clean image from the noisy input. In the self-supervised Noise2Noise framework lehtinen2018noise2noise, two noisy samples with identically distributed noise are required. For Noisier2Noise moran2020noisier2noise, the input sample is generated by further corrupting the given noisy sample with additional noise. For Noise2Self batson2019noise2self, the input image is restricted to a pre-defined mask and the estimated output is compared against the same input image projected on the complement of the employed mask. The principle of the framework self-supervision by data undersampling (SSDU) yaman2020self (not shown in this Figure) that can be used for MR image reconstruction is similar to Noise2Self with the difference that the mask is applied in the Fourier domain of the considered image.
• Figure 3: An example of images reconstructed with two deep neural networks for image denoising (Sparse Net and Joint Sparse-Net) that were trained using the Noise2Noise (N2N) framework on low-field MR image data acquired at $0.5\,\mathrm{T}$. The obtained images show a strong noise reduction compared to the images reconstructed from one and six averaged acquisitions (1NSA and NSA6), respectively. The image was adapted from Lian2022ismrm and reproduced with the permission of the ISMRM. Image courtesy of Hua Guo.
• Figure 4: An example of cine images acquired at $0.55\,\mathrm{T}$ using spiral acquisitions during free breathing and reconstructed with $\ell_1$-ESPIRIT uecker2014espirit, a low-rank + sparsity-based method (L+S) otazo2015low, the proposed low-rank deep image prior (LR-DIP)Hamilton2023magma and a reference image of the same subject obtained at $1.5\,\mathrm{T}$ acquired during a breathhold with an ECG-gated acquisition. The neural networks-based approach LR-DIP improves image quality over the two compressed sensing-based approaches and yields visually improved noise and artefacts suppression. The image was adapted from Hamilton2023magma. Image courtesy of Jesse Hamilton.
• Figure 5: An example of ultra-low-field ($6.5\, \mathrm{mT}$) MR images of the brain reconstructed with different methods. From left to right: AUTOMAP zhu2018imagekoonjoo2021boosting, simple inverse FFT, inverse FFT image denoised with DnCNN zhang2017beyond, inverse FFT image denoised with block-matching 3D dabov2006imagemakinen2020collaborative. Figure adapted from koonjoo2021boosting. Image courtesy of Neha Koonjoo.