Brain-age in ultra-low-field MRI: How well does it work?

Brain-age estimates the brain's biological age from neuroimaging data and has been proposed as a biomarker of brain health and disease risk. While brain-age estimation commonly uses high-field (HF) magnetic resonance imaging (MRI) ( $>1.5$ T), this is costly and inaccessible, limiting its applicability. Emerging ultra-low-field (ULF) MRI ( $<0.1$ T) is cheaper and more accessible, but its lower resolution may limit the reliability of biomarkers such as brain-age. We assessed different brain-age pipelines in 23 adults scanned on one HF system (GE Signa Premier at 3 T) and two identical ULF systems (Hyperfine Swoop at 64 mT) located at two different sites, hereafter referred to as ULF1 and ULF2. We used 14 distinct acquisitions defined by T1- or T2-weighting, resolution, and preprocessing: raw anisotropic orientations (axial, coronal, sagittal), isotropic scans, and super-resolution derivatives from multi-resolution registration (MRR) and SynthSR. These inputs (a total of n = 573 scans) were analyzed with five brain-age software packages (BrainageR, SynthBA, MIDI, DeepBrainNet, PyBrainAge). Performance evaluation entailed validity (brain-age vs. actual age), correspondence (ULF brain-age vs. HF brain-age) and test-retest reliability (ULF1 brain-age vs. ULF2 brain-age). Overall, results were mixed across pipelines, although several ULF pipelines performed comparably to HF. The four best-performing combinations were SynthBA on T2 scans without SynthSR, MIDI on T2 scans without SynthSR, PyBrainAge on T1 scans with SynthSR and using FreeSurfer recon-all-clinical, and BrainageR on T1 scans with SynthSR. These showed moderate-to-strong validity ( $r=0.76$ -0.92, $R^2\hspace{0.17em}=0.54$ -0.64, Mean Absolute Error [MAE] $=6.7$ -8.21 years), moderate-to-strong correspondence to HF ( $r=0.84$ -0.93, Intraclass Correlation Coefficient [ICC] $=0.72$ -0.92), and excellent test-retest reliability ( $r=0.97$ -0.99, ICC $=0.97$ -0.99). Moreover, some anisotropic acquisitions achieved comparable validity and reliability to MRR images when tested with the best-performing model, SynthBA ( $R^2\hspace{0.17em}=0.57$ -0.62, ICC [CI] $=0.99$ [0.97-1.00], for coronal T2). This first systematic evaluation of brain-age at ULF demonstrates that accurate and reliable estimates can be achieved across multiple pipelines, without necessarily requiring image enhancement. Performance depended on the combination of model, scan type, and preprocessing. ULF brain-age estimation could be a practical and scalable tool for clinical decision-making, population research, and long-term patient monitoring, thereby helping to make advanced neuroimaging biomarkers more accessible worldwide.