Brain-age estimation using ultra-low-field MRI (64 mT) can achieve moderate-to-strong validity, moderate-to-strong correspondence to high-field MRI, and excellent test-retest reliability across multiple pipelines, demonstrating that 'accurate and reliable estimates can be achieved across multiple pipelines, without necessarily requiring image enhancement.'
Key Findings
Results
The four best-performing ULF brain-age pipeline combinations demonstrated moderate-to-strong validity in estimating brain age from actual age.
Best-performing combinations were: SynthBA on T2 scans without SynthSR, MIDI on T2 scans without SynthSR, PyBrainAge on T1 scans with SynthSR using FreeSurfer recon-all-clinical, and BrainageR on T1 scans with SynthSR.
Validity metrics across these four combinations: r = 0.76–0.92, R² = 0.54–0.64, Mean Absolute Error (MAE) = 6.7–8.21 years.
Study included 23 adults scanned on one HF system (GE Signa Premier at 3 T) and two identical ULF systems (Hyperfine Swoop at 64 mT) at two different sites.
A total of n = 573 scans were analyzed across 14 distinct acquisitions and five brain-age software packages.
Results
The four best-performing ULF brain-age pipeline combinations showed moderate-to-strong correspondence to high-field MRI brain-age estimates.
Correspondence metrics (ULF brain-age vs. HF brain-age): r = 0.84–0.93, Intraclass Correlation Coefficient (ICC) = 0.72–0.92.
Correspondence was assessed by comparing ULF brain-age output to HF brain-age output from the same participants.
ICC values ranged from moderate (0.72) to strong (0.92) depending on the specific pipeline combination.
Results
The four best-performing ULF brain-age pipeline combinations demonstrated excellent test-retest reliability across two different ULF scanner sites.
Test-retest reliability metrics (ULF1 brain-age vs. ULF2 brain-age): r = 0.97–0.99, ICC = 0.97–0.99.
Test-retest reliability was evaluated using two identical Hyperfine Swoop 64 mT systems located at two different sites.
ICC values of 0.97–0.99 indicate excellent reproducibility of brain-age estimates across ULF sites.
Results
Some anisotropic (non-isotropic) ULF acquisitions achieved validity and reliability comparable to multi-resolution registration (MRR) enhanced images when used with the best-performing model SynthBA.
Coronal T2 anisotropic acquisitions achieved R² = 0.57–0.62 and ICC (CI) = 0.99 [0.97–1.00] with SynthBA.
This finding indicates that image enhancement through MRR is not always necessary to achieve reliable brain-age estimates.
The result applies specifically to the SynthBA model on T2-weighted coronal acquisitions.
Results
Overall results across all tested ULF brain-age pipeline combinations were mixed, with performance depending on the specific combination of model, scan type, and preprocessing.
Five brain-age software packages were tested: BrainageR, SynthBA, MIDI, DeepBrainNet, and PyBrainAge.
14 distinct acquisitions were defined by T1- or T2-weighting, resolution, and preprocessing including raw anisotropic orientations (axial, coronal, sagittal), isotropic scans, and super-resolution derivatives from MRR and SynthSR.
Not all combinations performed comparably to HF; only several ULF pipelines performed comparably.
The authors state that 'results were mixed across pipelines, although several ULF pipelines performed comparably to HF.'
Methods
Ultra-low-field MRI operates at 64 mT (Hyperfine Swoop), which is substantially lower field strength than the 3 T high-field system used for comparison, with lower resolution being a potential limitation for biomarker reliability.
ULF MRI is defined as < 0.1 T; the systems used were Hyperfine Swoop at 64 mT.
High-field MRI used for comparison was a GE Signa Premier at 3 T.
ULF MRI is described as 'cheaper and more accessible' than HF MRI, but its 'lower resolution may limit the reliability of biomarkers such as brain-age.'
The study enrolled 23 adults scanned across both modalities.
Discussion
The authors propose that ULF brain-age estimation could serve as a practical and scalable tool for clinical and population-level applications.
Proposed applications include 'clinical decision-making, population research, and long-term patient monitoring.'
The authors suggest ULF brain-age could 'help make advanced neuroimaging biomarkers more accessible worldwide.'
This was described as 'the first systematic evaluation of brain-age at ULF.'
What This Means
This research suggests that brain-age — a measure of how old the brain appears based on MRI scans, which can serve as a health biomarker — can be reliably estimated using ultra-low-field (ULF) MRI scanners operating at just 64 millitesla, far weaker than the standard 3 Tesla (3T) scanners used in hospitals and research. The study scanned 23 adults on one high-field 3T scanner and two identical ULF scanners at different locations, then tested five different brain-age analysis programs across 14 different scan types and image processing approaches. The best-performing combinations of ULF scanner settings and analysis software produced brain-age estimates that were reasonably accurate (off by about 6.7 to 8.2 years on average), strongly matched the estimates from the expensive high-field scanner, and were highly reproducible when the same person was scanned on two different ULF machines at different sites.
A notable finding was that not all approaches worked equally well — performance depended heavily on which analysis software, scan type (T1- or T2-weighted), and image processing steps were combined. However, the study also found that some simpler scan types, without additional image enhancement processing, performed just as well as more processed versions when paired with the right analysis software (specifically, SynthBA on T2-weighted coronal scans). This suggests that the added computational steps of image enhancement may not always be necessary.
This research matters because standard high-field MRI scanners are expensive, large, and unavailable in many parts of the world. Ultra-low-field MRI scanners are smaller, cheaper, and can potentially be used in community clinics or low-resource settings. This study suggests that these more accessible scanners could still provide meaningful brain health information through brain-age estimation, potentially enabling broader use of neuroimaging biomarkers in routine healthcare monitoring and large-scale population health research.
Biondo F, Bennallick C, Martin S, Puglisi L, Booth T, Wood D, et al.. (2026). Brain-age in ultra-low-field MRI: How well does it work?. Imaging neuroscience (Cambridge, Mass.). https://doi.org/10.1162/IMAG.a.1352