A team at the Cardiff University Brain Research Imaging Centre (CUBRIC) now reports what they describe as the first systematic test-retest evaluation of tract-based diffusion tensor MRI (DT-MRI) metrics on a portable 64 millitesla scanner. The study, published in Frontiers in Neuroimaging, demonstrates that despite the inherent signal limitations of ultra-low-field imaging, repeated measurements of key white-matter metrics are stable enough to support group-level statistical inference. The work is framed as a foundation for democratizing advanced neuroimaging and enabling population-scale brain studies in regions that have historically been excluded from the field.

Low-field MRI systems, defined by recent International Society for Magnetic Resonance in Medicine safety guidelines as operating below 0.1 tesla, use permanent magnets rather than superconducting ones. That makes them dramatically cheaper to buy, easier to install, free of cryogenic requirements, and genuinely portable thanks to a small footprint and low power demands. The trade-off is physics: at low field strengths, signal-to-noise ratios fall, forcing longer acquisitions and yielding images with lower resolution and contrast. These constraints have historically limited clinical and research applications, particularly for the signal-hungry quantitative diffusion metric maps and tractography that underpin modern brain connectivity research.

Recent advances in image reconstruction, acquisition strategies, and post-processing have begun to change that calculus. Prior work has already shown reasonable agreement between 3 tesla and 64 millitesla tractography. Building on those gains, the Cardiff team set out to answer a more demanding question: are the quantitative diffusion measurements themselves repeatable enough, scan after scan, to anchor the population-scale studies that global neuroscience increasingly demands?

Ten healthy participants, five men and five women aged between 18 and 32 with a mean age of 27.4 years, were scanned twice on a Hyperfine Swoop portable system, with a median interval of 13 days between sessions. The 70-minute protocol included a single non-diffusion-weighted reference volume, a 2-millimeter isotropic T2-weighted structural image, and 18 diffusion-weighted volumes acquired along different directions at a b-value of 900 s/mm² using a multi-shot fast-spin-echo sequence at 3-millimeter isotropic resolution. Short calibration scans were interleaved throughout the session to reset the scanner’s center frequency, compensating for drift in the main magnetic field as the scanner’s temperature rose during operation.

The processing pipeline was correspondingly sophisticated. Images were reconstructed with the vendor’s deep learning algorithm, then registered nonlinearly to the reference volume using the ANTs toolkit, with the resulting transformations also applied to rotate the diffusion-encoding vectors so that orientational information was preserved. A bias field correction addressed a previously reported directional bias caused by imperfections in the encoding gradient fields, and diffusion tensors were fitted in every brain voxel with a linear least-squares approach. The T2-weighted structural image was then super-resolved into a synthesized 1-millimeter T1-weighted image, segmented into tissue classes, and used to calibrate a single-fiber response function for constrained spherical deconvolution. The peaks of the resulting fiber orientation distributions fed TractSeg, a bundle-specific tractography method that segments anatomically defined white-matter tracts directly from orientation information without requiring whole-brain tractography as an intermediate step.

The team extracted tract-averaged fractional anisotropy (FA), mean diffusivity (MD), and radial diffusivity (RD) from a selection of major white-matter pathways spanning projection, commissural, and association fiber categories: the arcuate fasciculus, anterior thalamic radiation, two segments of the corpus callosum, the corticospinal tract, the cingulum, and the inferior fronto-occipital fasciculus. These tracts cover functions from motor control and language to thalamo-cortical connectivity, interhemispheric communication, and limbic cognition. Bland-Altman analysis and within-subject coefficients of variation indicated strong scan-rescan agreement, with virtually all data points falling within the limits of agreement. The few outliers for MD and RD all traced back to a single participant whose scan had abnormally high diffusivity values.

Reliability, measured by intraclass correlation coefficients, was more variable. For MD and RD, all tracts showed poor to moderate ICC values, while FA achieved good reliability in the arcuate fasciculus bilaterally and in both corpus callosum segments. The authors attribute the generally low ICC scores to their young cohort exhibiting minimal true inter-subject variability, combined with cerebrospinal fluid contamination pulling values toward a common mean, so that between-subject variation ended up smaller than measurement variance. Wide confidence intervals also indicated the sample of ten was too small for stable ICC estimation. After false discovery rate correction for multiple comparisons, only FA in the corpus callosum and arcuate fasciculus retained statistical significance.

Crucially, the team translated their repeatability figures into practical power calculations. Compared with high-field reference data, within-subject coefficients of variation were 3 to 10 times larger, and standard deviations in the group means were roughly an order of magnitude higher. Even so, the sample sizes required to detect group differences in a two-tailed t-test remained feasible: approximately 25 participants per group sufficed for detecting a 4 percent difference in means for the majority of tracts in MD and RD, and around 65 per group for FA. By comparison, at 3 tesla, detecting even a 1 percent difference in the corticospinal tract would require fewer than 10 participants per group. The authors caution that their MD and RD estimates may be optimistic, because cerebrospinal fluid contamination can artificially reduce inter-participant variability and inflate apparent statistical power.

The absolute metric values also diverged from high-field norms. FA in the corpus callosum splenium, the highest-value tract in the study, reached only 0.22, versus roughly 0.81 in healthy adults at 3 tesla, while MD values ranged from 1.05 to 1.2 × 10⁻³ mm²/s against literature values near 0.75 × 10⁻³ mm²/s in the corpus callosum. The authors attribute these discrepancies to the long echo train broadening the point spread function, causing tissue mixing and cerebrospinal fluid contamination; to partial volume effects from the large 3-millimeter voxels; to blurring introduced by nonlinear motion and distortion correction; and to a Rician noise-floor bias that compresses directional diffusivity differences and makes the diffusion tensor appear more isotropic. Smaller peripheral tracts, buried in noise, fared worst, though population-averaged maps recovered previously unresolved features including U fibers, the corona radiata, internal capsule, and cerebellar peduncles.

Looking forward, the authors see several routes to closing the gap. Sequence optimization and improved processing could shorten scan times and improve usability, particularly for clinical or developmental populations. Emerging approaches—including super-resolution methods, deep learning synthesis frameworks such as DeepDTI, and joint reconstruction techniques—exploit correlations across diffusion volumes to recover high angular resolution information from substantially reduced sampling schemes. Removing the need for the structural T2 image, by deriving response functions or white-matter masks differently, could trim the protocol further. The growing co-deployment of portable low-field and high-field scanners also creates an opportunity to accumulate paired datasets large enough to learn cross-field-strength signal representations, potentially using high-field priors to improve low-field image quality—though the authors note such models will likely generalize best within a shared scanner and protocol family rather than across heterogeneous low-field hardware. Repeatability studies of the kind presented here, they argue, will be an essential foundation for validating any such harmonization approach as the global low-field MRI footprint grows.

Subject of Research: Test-retest repeatability of tract-based diffusion tensor MRI metrics on a portable ultra-low-field 64 mT scanner

Article Title: Repeatable tract-based diffusion metrics on a portable 64 mT MRI: a foundation for global population neuroscience

Article References: Repeatable tract-based diffusion metrics on a portable 64 mT MRI: a foundation for global population neuroscience. (n.d.). https://doi.org/10.3389/fnimg.2026.1845201

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1845201

Keywords: low-field MRI, portable MRI, diffusion tensor imaging, tractography, repeatability, population neuroscience, white matter, fractional anisotropy, global health equity, 64 mT, test-retest, neuroimaging democratization