Low-Field MRI Consortium

• Low-field MRI Consortium is a collaborative framework that unites academic labs, hospitals, and companies to develop and benchmark low-field MRI systems with open-source tools.
• It integrates diverse scanner architectures and control systems, enabling reproducible hardware designs and seamless sequence interoperability across multiple platforms.
• Emerging techniques in super-resolution and metabolic imaging demonstrate the consortium's commitment to improving diagnostic accuracy in resource-limited and bedside settings.

The Low-Field MRI Consortium denotes a consortium-style research framework for developing, benchmarking, and disseminating low-field magnetic resonance imaging systems, associated control electronics, reconstruction software, and post-processing methods. In the literature summarized here, the concept is represented most explicitly by the open MaRCoS community, by workshop-based capacity-building models for constructing operational low-field scanners, and by multi-center methodological proposals for image enhancement, field monitoring, and metabolic imaging. The resulting ecosystem spans permanent-magnet and electromagnet platforms from millitesla to a few hundred millitesla, open-source console stacks, portable bedside neuroimaging, and algorithmic image-quality transfer intended to bridge the gap between low-field and high-field MRI (Guallart-Naval et al., 2022).

1. Consortium structure and collaborative models

A defining feature of low-field MRI consortia is the use of distributed, open collaboration rather than vertically integrated proprietary development. The MaRCoS consortium is described as an open, international collaboration of academic laboratories, spin-off companies, and research hospitals dedicated to developing a flexible, low-cost, high-performance console for low-field MRI. Its organizational structure is deliberately flat: a core steering committee manages overall direction and quality control, while working groups take responsibility for hardware design, firmware development, sequence programming, reconstruction algorithms, and user support. Governance is based on consensus and open-source licensing, and new groups are asked to contribute documentation and validation data, participate in monthly consortium meetings, and help maintain central repositories (Guallart-Naval et al., 2022).

A second collaborative model is the workshop-based consortium prototype demonstrated by the three-day “Delta DIY MRI” workshop at Johns Hopkins University. That workshop hosted 16 participants and organized them into four parallel subsystem teams: Magnet; Passive Shimming + Gradients; RF Coil + RF Shield; and Data Acquisition + Reconstruction. The workflow was milestone-driven, with daily check-ins and cross-team integration sessions; each morning began with a historical keynote contextualizing the day’s technical focus. Preparatory work included weekly planning meetings four months in advance, software setup sessions for PyPulseq and MaRCoS, and coordinated fabrication planning for CNC machining and 3D printing. Post-workshop recommendations included a Consortium Secretariat, shared repositories of CAD, G-code, and curricula, and “train the trainer” modules for regional scaling (Oiye et al., 26 Nov 2025).

Taken together, these reports present the consortium not merely as an administrative entity but as a technical coordination mechanism. A plausible implication is that low-field MRI progress depends on coupling governance, reproducibility, and education as tightly as it couples magnetics, RF engineering, and reconstruction.

2. Scanner architectures and hardware realizations

The hardware base of consortium-style low-field MRI spans markedly different geometries and field strengths. In the MaRCoS benchmarking study, three systems were highlighted: a 360 mT tabletop dipole scanner at Massachusetts General Hospital; a 50 mT human Halbach scanner with a 27 cm bore at Leiden University Medical Center; and a 70 mT Halbach system at i3M, which extended the LUMC design and improved homogeneity from 15 700 ppm to 3 100 ppm over a 20 cm DSV by adding 1 100 shim magnets. These systems also differed in gradient topology, RF hardware, and intended use, ranging from education and rapid prototyping to in vivo extremity imaging (Guallart-Naval et al., 2022).

The workshop-built 0.27 T scanner illustrates a compact, open-hardware realization. Its magnet used six N45 NdFeB discs assembled in two triplets, with AISI 1018 steel yoke and pole pieces and a 9 cm pole separation. The simulated center field was $B_{0,\text{sim}} = 0.277\,\mathrm{T}$, and the measured center field was $B_{0,\text{meas}} = 0.269\,\mathrm{T}$. Field inhomogeneity over a 3.2 cm DSV improved from $3\,\mathrm{mT}$ before shimming to $2\,\mathrm{mT}$ after passive shimming. The associated RF solenoid had 3 cm inner diameter, 3.4 cm outer diameter, and 6.8 cm length; it was tuned to 11.4 MHz and matched to $50\,\Omega$, with a reflection coefficient of –35 dB. Its planar gradient coils were generated via genetic algorithms and exhibited simulated non-linearity below 5% over a 3 cm FOV (Oiye et al., 26 Nov 2025).

Several compact analytical expressions appear in this hardware literature. For the H-magnet, a simplified relation is

$B_0 \approx \frac{\mu_0 M_r t}{g},$

while field uniformity is quantified by

$$\eta = \frac{\max(B)-\min(B)}{\overline{B}} \times 100\%.$$

These formulas are used as design heuristics rather than complete field models. More generally, the hardware record shows that low-field MRI is not tied to a single architecture: dipole, Halbach, H-shaped permanent-magnet, and open electromagnet configurations all appear in current consortium-oriented work.

3. Open-source control systems and sequence interoperability

Open control electronics and software are central to consortium operation because they permit multi-site benchmarking and rapid sequence transfer. The MaRCoS architecture is built around a Red Pitaya SDRlab board containing a Xilinx Zynq FPGA and ARM CPU, with fast ADCs, DACs, and GPIO. FPGA firmware known as flocra implements timing for RF pulses, gradients, digitization, and pre-processing; the ARM runs the MaRCoS server as a TCP/IP bridge; and the host computer runs a Python client with either PhysioMRI_GUI or a PulSeq interpreter. Pulse sequences may be entered through the native Python GUI, prepared externally as PulSeq .seq files in MATLAB or Python, or directly coded as Numpy arrays or text files. After each repetition, the FPGA streams pre-processed k-space data back to the client, where user-defined FFT or NUFFT reconstruction is invoked (Guallart-Naval et al., 2022).

The implemented sequence set is already broader than simple spin echo. The benchmarking paper lists Spin Echo, Rapid Acquisition with Refocused Echoes, Gradient Echo, Short Tau Inversion Recovery, and radial non-Cartesian readouts. It also gives the familiar low-field signal model

$$S(\rho,T_1,T_2) \propto \rho \,\bigl(1 - e^{-TR/T_1}\bigr)e^{-TE/T_2}\sin \alpha,$$

and notes that $\mathrm{SNR} \propto B_0^p$ with $p \approx 1$–1.5 in low-field regimes. In the i3M system, 3D-RARE reduced scan time from 15 min to 1.5 min, while LUMC reported a 30% SNR boost from parallel reception (Guallart-Naval et al., 2022).

The workshop system used the same open ecosystem in a more pedagogical configuration: MaRCoS firmware on a Red Pitaya SDRlab 122-16, MRI4ALL software on Ubuntu 22.04, and a FLOCRA-Pulseq bridge from PyPulseq to MaRCoS. The reconstruction chain was explicit—raw ADC data to pre-filtering to FFT to image space to magnitude extraction—and was validated on phantom studies, including SNR measurements in water and CuSO$B_{0,\text{meas}} = 0.269\,\mathrm{T}$0 regions and spatial-resolution checks on partitioned phantoms (Oiye et al., 26 Nov 2025).

This control-layer standardization has broader significance. It makes sequence programming, hardware replacement, and benchmarking portable across scanner designs, which is a necessary condition for any consortium attempting reproducible low-field MRI across sites.

4. Image enhancement, super-resolution, and morphometry

Low-field MRI’s most immediate technical limitation is degraded spatial resolution and SNR relative to conventional high-field systems. Portable 64 mT scanners allow bedside neuroimaging of critically ill patients and avoid transport risks, but their SNR is approximately $B_{0,\text{meas}} = 0.269\,\mathrm{T}$1 lower and voxel sizes are typically about $B_{0,\text{meas}} = 0.269\,\mathrm{T}$2, versus $B_{0,\text{meas}} = 0.269\,\mathrm{T}$3 isotropic in 1.5–3 T MRI. This gap prevents direct use of standard morphometry tools such as FreeSurfer, FSL, and SPM. In paired LF/HF data, direct application of automated segmentation to raw LF-MRI yielded mostly failed or uncorrelated results, whereas post–super-resolution processing recovered substantial volumetric correspondence with high-field measurements: hippocampus $B_{0,\text{meas}} = 0.269\,\mathrm{T}$4, thalamus $B_{0,\text{meas}} = 0.269\,\mathrm{T}$5, whole cerebrum $B_{0,\text{meas}} = 0.269\,\mathrm{T}$6, ventricles $B_{0,\text{meas}} = 0.269\,\mathrm{T}$7, and volumetric bias below 5% for large structures (Iglesias et al., 2022).

The super-resolution method extends the SynthSR framework with a segmentation-based regularizer. A synthetic data generator begins from 20 high-field MPRAGE scans with 39 ROI labels, applies random nonlinear deformation, samples synthetic T1/T2 intensities through a bivariate Gaussian mixture model conditioned on labels, and degrades these to low-field characteristics by downsampling, Rician noise, bias field corruption, and resampling. A five-level 3D U-net then maps two input channels—synthetic LF-T1 and LF-T2—to a 1 mm isotropic synthetic MPRAGE. The joint loss is

$B_{0,\text{meas}} = 0.269\,\mathrm{T}$8

with $B_{0,\text{meas}} = 0.269\,\mathrm{T}$9 enforcing intensity fidelity and $3\,\mathrm{mT}$0 enforcing segmentation fidelity through Dice overlap with a frozen segmentation U-net (Iglesias et al., 2022).

A later image-quality transfer approach uses Conditional Flow Matching rather than a super-resolution loss tied to segmentation. In that framework, the target image is generated through the ODE

$3\,\mathrm{mT}$1

where the network regresses a continuous velocity field conditioned on the low-field input. The architecture is a modified U-Net with multi-scale convolutions, residual blocks with GroupNorm and SiLU, time embeddings, Squeeze-and-Excitation, pixel-unshuffle/pixel-shuffle resampling, and a transformer bottleneck. On 404 paired images derived from Human Connectome Project 3 T scans and simulated low-field degradations, IQT-CFM achieved the highest PSNR, SSIM, and lowest LPIPS on both in-distribution and out-of-distribution test sets while using approximately 5.25 million parameters, compared with 9.34 million for the IQT-DDL comparator. The reported in-distribution PSNR values were 28.92 dB for interpolation, 36.07 dB for IQT-DDL, and 37.07 dB for IQT-CFM; corresponding OOD PSNR values were 24.69 dB, 26.25 dB, and 26.33 dB (Nguyen et al., 14 Oct 2025).

These studies treat low-field reconstruction as more than denoising. They specifically target interoperability with morphometry pipelines, robustness to domain shift, and deployability in resource-limited settings. At the same time, both reports identify unresolved issues: small paired validation sets, single-scanner dependence, missing uncertainty estimates, and the need for sequence and device generalization.

5. Field monitoring, shimming, and quality control instrumentation

Consortium-scale deployment requires reproducible field metrology, because low-field image quality is tightly coupled to $3\,\mathrm{mT}$2 homogeneity and stability. A recent open NMR field probe addresses this by using a strong pre-polarization field $3\,\mathrm{mT}$3, up to 70 mT, applied perpendicular to $3\,\mathrm{mT}$4, followed by rapid non-adiabatic switch-off via silicon carbide transistors operated in controlled avalanche breakdown. The switch-off violates the adiabatic condition and leaves a transverse component

$3\,\mathrm{mT}$5

which then precesses at

$3\,\mathrm{mT}$6

Because $3\,\mathrm{mT}$7 is inferred from the free-induction-decay frequency, no prior field knowledge or sweep is required (Povolni et al., 12 Jan 2026).

The reported sensor spans 1 mT to 45 mT, with a typical sequence time of roughly 10 s per measurement point and a coarse mapping protocol using 10 mm steps to reproduce 3D field maps in approximately 10 min. Example no-averaging SNR values were about 21 dB at $3\,\mathrm{mT}$8, 35 dB at 3 mT, 38 dB at 5 mT, and 25 dB at 45 mT. The probe was used to identify $3\,\mathrm{mT}$9 drift in a Halbach magnet, to resolve $2\,\mathrm{mT}$0 at 3 mT in an ultra-low-field solenoid, and to verify X/Y/Z gradient fields with DC sweeps of $2\,\mathrm{mT}$1. Its modular bill of materials was reported as approximately $2\,\mathrm{mT}$2, excluding the shared oscilloscope, and hardware files were released in an open repository (Povolni et al., 12 Jan 2026).

This instrumentation complements the passive shimming and gradient-design methods seen elsewhere in the consortium literature. The workshop scanner used 3D-printed shim trays with 0.6 cm$2\,\mathrm{mT}$3 NdFeB magnets, and its genetic algorithm minimized

$2\,\mathrm{mT}$4

A plausible implication is that consortium standardization will increasingly depend on a full calibration stack: simulation, passive shimming, field mapping, drift monitoring, and possibly closed-loop correction.

6. Emerging directions: metabolic imaging, scaling, and unresolved limitations

Low-field MRI consortia are no longer confined to structural proton imaging. One notable extension is ultra-low-field $2\,\mathrm{mT}$5C MRI of hyperpolarized pyruvate at $2\,\mathrm{mT}$6, combining an open-access scanner with in situ SLIC SABRE hyperpolarization. In that system, a biplanar electromagnet was shimmed to achieve a $2\,\mathrm{mT}$7H water linewidth below 0.5 Hz, field-frequency lock maintained resonance within $2\,\mathrm{mT}$8 Hz, and a two-layer solenoid tuned to 69.5 kHz detected $2\,\mathrm{mT}$9C. The method achieved $50\,\Omega$0C signal enhancements several million times above thermal equilibrium, corresponding to polarization levels of approximately 3%, and enabled both J-coupling-resolved spectroscopy and 3D imaging without field cycling or sample shuttling (Boele et al., 12 Sep 2025).

The metabolic imaging report also frames its significance at consortium scale. It describes pathways for external parahydrogen polarization, rapid transfer, automation of gas-handling and temperature-control circuits, and adoption of open-source designs across low-field MRI networks. The projected hardware simplification is explicit: ultra-low-field MRI scanners are estimated below $50\,\Omega$110 k with rapid sample turnover below 10 s (Boele et al., 12 Sep 2025).

Across the literature, future directions recur with notable consistency. The MaRCoS roadmap lists steady-state free precession, zero echo time, spiral imaging, model-based and deep-learning reconstruction, compressed sensing, more receive channels, and real-time feedback control (Guallart-Naval et al., 2022). The super-resolution work calls for multi-site validation across different 64 mT systems, extension to FLAIR, diffusion, and T2$50\,\Omega$2, inference optimization for near-instant feedback, and Bayesian or Monte Carlo dropout uncertainty estimation (Iglesias et al., 2022). The Conditional Flow Matching study points to common calibration protocols, local fine-tuning, privacy-preserving delta-model exchange, and mitigation of artifacts under extreme noise regimes with $50\,\Omega$3 (Nguyen et al., 14 Oct 2025). The field-probe study proposes arrayed sensors, integrated FPGA/ADC boards, and embedding probe measurements within image acquisition for live $50\,\Omega$4 correction (Povolni et al., 12 Jan 2026).

A persistent misconception is that low-field MRI is defined solely by reduced magnet strength. The published record summarized here suggests a broader and more accurate formulation: low-field MRI consortium work is a systems problem combining magnet design, open consoles, sequence portability, field control, reconstruction, image-quality transfer, and distributed training. In that sense, the consortium is both a technical architecture and an institutional one.