Population-based Brain Templates for Ultra-Low-Field MRI

Abstract

Ultra-low-field (ULF) MRI currently lacks population-representative spatial priors, limiting robust alignment, normalization, and cross-study comparability. Our aim is to provide an open, standardized ULF brain template resource, comprising data and code, that enables reproducible spatial analysis and method development across ULF studies. We present group-average brain templates generated from 64 mT MRI scans of 100 healthy adults, encompassing both T1- and T2-weighted contrasts and covering the full adult lifespan. Participants were stratified into three age groups to support both age-specific and population-level analyses. The preprocessing and registration pipeline employed established open-source neuroimaging tools and iterative averaging to enhance anatomical correspondence and consistency. Data were acquired across multiple scanner software versions, and potential intensity variability was evaluated and mitigated through standardized preprocessing. The resource includes population and age-specific templates with accompanying example segmentations and complete scripts for full reproducibility. Derived template data and code are openly available on Zenodo and GitHub in accordance with FAIR principles. This resource provides ULF-specific spatial priors to support normalization, registration benchmarking, and method development within comparable acquisition settings.

Background & Summary

Ultra-low-field (ULF) magnetic resonance imaging (MRI) systems, such as 64 millitesla (mT) portable scanners, are rapidly emerging as cost-effective and accessible alternatives to conventional high-field (1.5 T or 3 T) MRI, particularly in resource-limited, rural, or point-of-care (PoC) clinical environments^1–6. However, despite their increasing utility, ULF MRI systems produce images with lower spatial resolution and signal-to-noise ratio (SNR), which limits their direct integration into conventional image-processing pipelines or comparative neuroimaging studies^7–9. Since population-based brain templates, such as MNI/ICBM152 and its ICBM2009c variants for adults, and the NIHPD pediatric series, underpin normalization, segmentation, and cross-cohort comparison, the absence of ULF-specific templates matched to ULF contrast and resolution poses a critical barrier^10–15. To bridge this gap and support the development of advanced image-enhancement, segmentation, and harmonization techniques, there is an urgent need for population-representative brain templates derived directly from ULF MRI acquisitions.

Beyond hardware limitations, a fundamental challenge in ULF neuroimaging lies in the absence of standardized spatial priors and age-appropriate atlases that enable accurate morphometric analysis. Conventional brain templates such as the ICBM152 and MNI series were derived from high-field datasets^10,11, and direct application to ULF data risks systematic misregistration and bias due to contrast and resolution disparities. Deformation-based and voxel-based morphometry frameworks have demonstrated the value of population-specific templates for capturing age- and disease-related structural variability in high-field imaging^16,17. Extending these principles to the ULF domain is therefore essential for enabling unbiased structural quantification and cross-field harmonization.

Here, to our knowledge, we present the first open-access, population-representative brain template resource constructed exclusively from ULF (64 mT) MRI acquisitions. Built from T1- and T2-weighted scans of healthy adults spanning the adult lifespan, this resource provides standardized reference spaces tailored to ULF image contrast and resolution. The inclusion of both T1- and T2-weighted templates supports multimodal alignment and broadens applicability for atlas-based analyses, registration benchmarking, and method development within the ULF domain. Rather than replacing subject-level datasets, the present resource is intended to provide ULF-specific spatial priors that can support normalization, population-level reference mapping, and reproducible evaluation of downstream processing methods. All scans were obtained using a commercially available, Food and Drug Administration and Health Canada-approved 64 mT PoC MRI system, Hyperfine Swoop (Hyperfine Research, Guilford, CT)¹⁸. By encompassing a broad age range, the dataset enables the study of age-related variation in brain morphology and supports the development of both age-specific and population-averaged templates, foundational tools for advancing harmonization, segmentation, and enhancement techniques in this emerging imaging domain¹⁹.

To ensure reproducibility and compatibility with standard neuroimaging pipelines, all DICOM images were converted to NIfTI format using the open-source dicom2nifti Python library and subsequently processed through a uniform preprocessing workflow. Intensity inhomogeneity was corrected using FSL FAST²⁰, and skull stripping was performed using FSL’s Brain Extraction Tool (BET)²¹. Linear registration to the initial template space was conducted using ANTs with the AffineFast transformation. Nonlinear registration was subsequently performed using the Symmetric Normalization (SyN) algorithm implemented in ANTs, allowing for high-precision alignment through multiresolution optimization. Final population and age-specific templates were generated via iterative averaging to enhance anatomical correspondence and minimize subject-specific variability.

The resulting templates also provide a foundation for advanced morphometric analyses, including deformation-based assessment of age-related structural changes¹⁷ and cross-modal harmonization between ULF and high-field MRI via deep learning-based registration and synthesis methods²². By integrating open-access template data, preprocessing code, and quality-control documentation, implemented using freely available, open-source neuroimaging tools such as FSL and ANTs, in accordance with FAIR data principles, this dataset aims to establish a standardized, reproducible reference for future ULF MRI research²³. We note that subject-matched ULF and high-field MRI datasets provide complementary value for cross-field validation and harmonization²⁴. In contrast, the present resource focuses on providing ULF-specific population and age-stratified spatial priors to support normalization, registration benchmarking, and atlas-based analyses directly within the ULF domain. While ULF MRI expands access to neuroimaging in point-of-care and resource-limited settings, it does not currently match high-field MRI in spatial resolution or contrast richness and should be viewed as complementary rather than equivalent. Furthermore, ULF image appearance and geometric fidelity can be scanner- and protocol-dependent; these templates should be interpreted as a standardized reference for the acquisition configuration described here, with the accompanying code intended to facilitate re-derivation on other ULF systems and protocols. We emphasize that the present resource is not intended to replace subject-level datasets, but rather to provide standardized spatial reference spaces for ULF MRI. While subject-level datasets enable statistical modeling, machine learning, and cross-field validation, template-based resources provide complementary value by supporting spatial normalization, atlas construction, and reproducible benchmarking across studies.

Methods

Participants and Dataset Composition

The dataset consists of T1- and T2-weighted ULF (64 mT) MRI scans from n = 100 healthy adult participants. Given the well-documented morphological changes that occur in the human brain over the lifespan^25–27, including significant alterations in gray matter, white matter, and cerebrospinal fluid volumes, the use of a single, non-age-specific template can introduce substantial anatomical bias and lead to inaccurate normalization and analysis of individual brains^10,28. To address this, we have adopted a method consistent with established neuroimaging practices by creating three age-specific brain templates corresponding to the age groups in our study: young, 18–40 years; middle, 41–60 years; and older age, 61–80 years²⁹. This approach is justified by foundational research in brain atlas creation, which has demonstrated that age-appropriate templates improve the accuracy of spatial normalization and tissue segmentation^30,31. By minimizing the anatomical distance between individual subjects’ brains and their respective group templates, we can ensure that our analysis of structural changes is not confounded by developmental or aging-related volumetric differences, thereby enhancing the precision and scientific validity of our findings. To fully utilize the 100 available samples and account for a subject with a chronological age of 81 years, we extended the upper limit of the older age group to include this individual (Table 1). All participants provided informed consent, and inclusion criteria required the absence of any known neurological or psychiatric disorders. The study protocol was approved by the Monash University Human Research Ethics Committee (approval no. 32631). This study’s procedures and image acquisitions were performed in accordance with the relevant guidelines, regulations, and institutional ethics requirements.

Table 1.

Primary analytic age groups used in this study.

Age group	N	Male, n (%)	Female, n (%)	Mean  ±  SD	Median [IQR]
Young (Y) 18-40	47	20 (42.6%)	27 (57.4%)	29.4  ±  6.8	30.0 [22.0-35.0]
Middle (M) 41-60	27	12 (44.4%)	15 (55.6%)	51.7  ±  5.9	52.0 [47.5-57.0]
Old (O) 61-81	26	12 (46.2%)	14 (53.8%)	70.2  ±  5.5	70.5 [65.2-74.8]

MRI Acquisition and Scanner Hardware, Software, and Image Reconstruction

All scans were acquired on the Hyperfine Swoop^® 64 mT portable MRI system (Hyperfine Research, Guilford, CT), a permanent-magnet ultra-low-field scanner designed for point-of-care neuroimaging. The system employs a 64 mT permanent magnet without cryogenic cooling and uses an 8-channel receive/1-channel transmit head coil integrated into the scanner housing. Image acquisition and system operation were controlled via a vendor-provided touchscreen interface.

Data collection for this study spanned three vendor software versions identified from the DICOM Software Versions field (0018,1020): 8.4.1, 8.5.1, and 8.6.0. All T1- and T2-weighted sequences were acquired using the same protocol definitions across these versions. Potential version-dependent differences in image intensity characteristics could not be excluded a priori and were therefore evaluated explicitly in the Technical Validation section.

Image reconstruction was performed entirely using the vendor’s proprietary reconstruction pipeline prior to data export. The research team had no access to or control over internal reconstruction algorithms or acquisition parameter modifications beyond the standard clinical protocol settings. All data used in this study were obtained as reconstructed DICOM images following completion of the full vendor reconstruction process.

For all sequences, the Swoop system employs an advanced deep learning-based reconstruction framework comprising three components: advanced gridding, advanced denoising, and advanced interpolation. These models are applied during image reconstruction to reduce noise, blurring, and pixelation and to improve overall image uniformity at ultra-low field strength. The deep learning reconstruction models are not trained to remove acquisition-related artifacts (e.g., motion or external RF interference). Each model incorporates internal data-consistency checks; if a reconstructed output is deemed inconsistent with the input data, the system automatically bypasses the deep learning component and falls back to a conventional linear reconstruction stage. In such cases, the DICOM Reconstruction Algorithm tag is set to Linear for T1- and T2-weighted images, as described in the manufacturer’s technical documentation³². No additional deep learning-based post hoc upsampling or enhancement was applied beyond the vendor’s standard reconstruction pipeline. All subsequent processing steps described in this manuscript were performed on these reconstructed images.

For the T1-weighted sequence, an axial spin-echo inversion recovery (SE/IR) protocol was used with TR = 1500 ms, TE = 6.0 ms, TI = 300.0 ms, matrix size = 112 × 138, and slice thickness = 5.0 mm, and scan duration of 6.17 minutes. For the T2-weighted sequence, an axial spin-echo protocol was used with TR = 2000 ms, TE = 203.6 ms, matrix size = 120 × 146, and slice thickness = 5.0 mm, and scan duration of 6.30 minutes. Both protocols were optimized for gray/white matter contrast at 64 mT field strength. No contrast agent was administered. T1- and T2-weighted images were acquired using the same slice orientation to facilitate robust cross-contrast registration and template construction, and to minimize interpolation-induced inconsistencies in multimodal alignment. While rotated fields-of-view can improve effective resolution in some scenarios, consistent orientation was prioritized here to support stable groupwise averaging in ULF data with anisotropic slice thickness.

Preprocessing and Skull-Stripping

Raw DICOM datasets were first converted to NIfTI format using dicom2nifti to ensure compatibility with downstream processing pipelines. Following conversion, intensity inhomogeneities were corrected using the bias field correction module of FSL’s FAST (Fig. 1), which applies a hidden Markov random field model with an expectation-maximization algorithm to estimate and remove low-frequency intensity bias²⁰. This step substantially improved signal uniformity across the brain volume, particularly in ULF MRI scans prone to coil sensitivity variations. Subsequently, skull stripping was performed using FSL’s Brain Extraction Tool (BET) with an empirically optimized fractional intensity threshold of 0.4²¹. Representative outcomes of BET skull stripping, including masks and overlays, are shown in Fig. 2. No subjects were excluded during preprocessing; all brain masks were visually inspected to confirm accurate delineation of intracranial structures and exclusion of non-brain tissues. The bias-corrected, skull-stripped images were then used in subsequent registration and template-building stages to ensure consistent anatomical representation across participants.

Fig. 1.

Qualitative and quantitative evaluation of bias field correction on ULF T1- and T2-weighted MRI scans. Top row: Mid-sagittal, mid-coronal, and mid-axial slices from original and bias-corrected T1- and T2-weighted images (left-right: original vs. bias-corrected). Middle row: Intensity profiles (middle rows of each pair) extracted along the horizontal midline from averaged views (n = 5 slices) illustrate improved signal homogeneity after correction. Bottom row: The same anatomical planes are shown after resampling to 1 mm isotropic resolution, demonstrating consistency of spatial integrity post-correction. These visualizations collectively highlight the effectiveness of bias field correction in improving image uniformity and anatomical contrast in ULF MRI.

Fig. 2.

Visualization of skull stripping on ULF brain MRI. Orthogonal slices (sagittal, coronal, axial) are shown for two representative subjects in T1- and T2-weighted contrasts. Rows depict (top) original images with brain mask contours overlaid, (middle) binary brain masks, and (bottom) brain-extracted images with contours for reference. Robust percentile-based intensity normalization (1st-99th percentile) was applied for consistency across modalities. The overlaid contours highlight accurate delineation of brain boundaries while preserving gray-white matter contrast and minimizing non-brain tissue.

Isotropic Resampling

To ensure consistency in spatial resolution across all volumetric brain MRI datasets, we employed a standardized resampling procedure that converts each image to an isotropic voxel size of 1 × 1 × 1 mm³. This process involves estimating the appropriate scaling factors from the original voxel dimensions and applying trilinear interpolation to generate a resampled volume with uniform spatial resolution. The associated affine transformation matrix is concurrently updated to preserve anatomical orientation and spatial correspondence. This step is critical for enabling accurate downstream analysis, such as intensity normalization, volumetric comparison, and deep learning-based enhancement or segmentation, by mitigating the effects of voxel anisotropy and scanner-specific acquisition parameters³³.

Linear Registration

Linear (affine) registration was performed using the Advanced Normalization Tools (ANTs) library implemented in Python via the ANTs package. Given the relatively modest sample size in our cohort, ANTs was selected for its proven ability to achieve stable and anatomically consistent population averages even when fewer subjects are available³⁴. Each subject’s bias-corrected and skull-stripped ULF scan was aligned to a fixed T2-weighted reference image configuration (AffineFast). This transformation applies a 12-parameter affine model optimized through mutual information similarity and multi-resolution iterative refinement, providing an efficient yet robust alignment suitable for large datasets. All registered outputs were saved as spatially aligned volumes, and their affine transformation parameters were retained to ensure reproducibility and facilitate subsequent nonlinear registration. The resulting linearly aligned images were averaged to generate an initial group-level template for each modality. This template served as the reference target for the following nonlinear registration and iterative template refinement procedures. The progressive sharpening and convergence of the group template across the affine and subsequent nonlinear iterations are summarized in Fig. 3.

Fig. 3.

Comparison of T1- and T2-weighted low-field MRI templates across linear and nonlinear registration iterations. Each row corresponds to one template iteration, beginning with the linearly registered template and followed by five nonlinear registration refinements. For each modality (T1 and T2), the mid-sagittal, mid-coronal, and mid-axial slices of the group template are shown alongside their respective voxel-wise intensity difference maps relative to the preceding iteration (I_k − I_k−1). Before differencing, all volumes were resampled to 1 mm isotropic resolution and intensity-normalized using robust percentile scaling (1st–99th percentile mapped to [0, 1]). Difference maps are displayed using a zero-centered diverging blue–white–red colormap: blue indicates a decrease (I_k < I_k−1), red indicates an increase (I_k > I_k−1), and white represents minimal or no change. The color scale is symmetric and fixed across all panels to the 99th percentile of ∣I_k − I_k−1∣ pooled across all slices (arbitrary normalized intensity units). Progressive iterations show reduced magnitude and spatial extent of differences, reflecting improved anatomical correspondence and convergence of the group template.

Nonlinear Registration

Nonlinear image registration was performed using the ANTs software, employing the antsRegistrationSyNQuick[s] transformation type, which implements the Symmetric Normalization (SyN) algorithm in a streamlined configuration optimized for structural MRI alignment. Each participant’s linearly aligned T1-weighted image was nonlinearly registered to the initial group-average template. The procedure used default multi-resolution settings with symmetric diffeomorphic deformation and cross-correlation (CC) as the similarity metric. The resulting affine and nonlinear warp fields were saved along with the warped anatomical outputs, while inverse transforms were retained to enable template-to-native space mapping. All registrations were executed in batch mode for consistency, and the outputs were organized systematically to ensure reproducibility. The nonlinearly aligned scans were subsequently used for iterative template refinement and cohort-level averaging. Age-stratified templates and their voxel-wise difference maps following nonlinear alignment are presented in Fig. 4.

Fig. 4.

Orthogonal mid-slice views (sagittal, coronal, and axial) of the group-wise average brain templates constructed from ULF MRI scans for both T1- and T2-weighted sequences, along with corresponding voxel-wise difference maps between consecutive age groups. Each row represents a different age group (Young, Middle, and Old). For each modality, anatomical views are shown in grayscale and immediately followed by their difference maps (shown in red-blue diverging colormap). Columns are arranged as follows: T1 sagittal, T1 sagittal difference, T1 coronal, T1 coronal difference, T1 axial, T1 axial difference, followed by the same layout for T2. All images were resampled to isotropic resolution for consistent comparison.

Inter-subject Variability After Registration

Inter-subject variability was quantified separately for T1- and T2-weighted images using voxel-wise coefficient-of-variation (CoV) maps computed across subjects in template space following robust intensity normalization. For T1-weighted images, the median CoV within the brain mask decreased monotonically from 0.307 after affine-only alignment to 0.084 after the final nonlinear iteration, corresponding to a reduction of approximately 73%. The 95th percentile CoV similarly decreased from 6.84 to 2.63, indicating reduced residual variability in high-variance regions such as CSF boundaries and cortical interfaces. For T2-weighted images, a consistent pattern was observed, with median CoV decreasing from 0.255 after affine alignment to 0.118 after the final nonlinear iteration (approximately 54% reduction), and the 95th percentile CoV decreasing from 7.04 to 2.53. Across both contrasts, mean CoV values also showed progressive reductions with successive nonlinear iterations. Together, these results demonstrate systematic and modality-consistent reductions in inter-subject variability, reflecting improved anatomical correspondence with nonlinear registration and supporting the robustness of the ULF template construction pipeline for both T1- and T2-weighted data.

Template Construction

Template construction was performed iteratively. Initially, a randomly selected subject from the dataset served as the starting reference (iteration 0). After linear and nonlinear alignment of all other subjects to this reference, their aligned volumes were averaged voxel-wise to produce an updated group mean template. This iterative registration and averaging process was repeated until visual convergence (5 iterations). Intermediate and final templates were visually assessed for image sharpness, anatomical definition, and overall symmetry to ensure consistent alignment and structural fidelity. Age-specific subgroup templates were also constructed by dividing the cohort into three subsets (Table 1) to evaluate the robustness and age-related consistency of the template. No Laplacian, high-pass, edge-enhancing, or unsharp masking filters were applied at any stage of template construction. The observed increase in anatomical definition across iterations arises solely from improved nonlinear alignment and iterative voxel-wise averaging.

Interpretation of Age-Related Volumetric Trends

These age-stratified templates are intended as descriptive, population-level anatomical priors and are not designed to replace subject-level analyses or regression-based modeling of aging trajectories. Table 2 summarizes brain region (combined left and right hemisphere) volumes across three age groups (Young: 18–40 years, Middle: 41–60 years, Old: 61–81 years) derived from T1-weighted group-average templates. The data reveal heterogeneous and region-specific volumetric trajectories across the adult lifespan. Segmented T1 and T2 group templates across age groups, highlighting regional anatomy and consistency, are visualized in Fig. 5. Segmentation was performed on the group-average templates to demonstrate the anatomical usability of the released reference spaces and to provide descriptive regional summaries. These template-level segmentations are not intended for statistical inference or replacement of subject-level segmentation and regression-based analyses.

Table 2.

Brain Region Volumes (in cm³) Across Age Groups.

Region	Young	Middle	Old
3rd ventricle	1.607	1.647	1.678
4th ventricle	3.408	3.708	3.39
Accumbens area	1.166	1.107	1.174
Amygdala	3.791	3.666	3.755
Brain-stem	27.977	27.643	28.045
CSF	296.369	302.939	298.655
Caudate	8.673	8.463	8.673
Cerebellum cortex	108.959	108.366	109.665
Cerebellum white matter	39.141	38.537	39.286
Cerebral cortex	533.829	527.789	531.049
Cerebral white matter	547.476	548.542	546.309
Hippocampus	9.103	8.861	9.066
Inferior lateral ventricle	1.275	1.525	1.52
Lateral ventricle	32.765	35.949	33.57
Pallidum	3.538	3.401	3.607
Putamen	12.894	12.639	13.076
Thalamus	17.11	16.66	17.108
Ventral DC	9.611	9.363	9.647

Fig. 5.

Representative orthogonal slices from segmented T1- and T2-weighted group-average brain templates across three age groups (Young (Y) 18–40, Middle (M) 41–60, Old (O) 61–81). For each age group (rows), sagittal, coronal, and axial views are shown for both T1 (left columns) and T2 (right columns) templates. Segmentations were performed using SynthSeg⁺³³ and harmonized to a common anatomical label set by merging corresponding left and right hemisphere structures. Distinct brain regions are color-coded and overlaid on the respective templates. This visualization provides qualitative evidence of structural differences across aging and modalities, highlighting both tissue-specific contrast and spatial consistency achieved by the groupwise template construction pipeline.

As expected, ventricular and CSF compartments demonstrated age-related enlargement, particularly from young to middle age. The lateral ventricles increased by +9.7% and the 4th ventricle by +8.8% between the young and middle-aged groups, while total CSF volume rose slightly (+2.2%). In older adults, ventricular volumes showed partial normalization, consistent with inter-subject variability and nonlinear aging trajectories. These findings are consistent with well-established ventricular expansion accompanying normal brain aging and parenchymal atrophy reported in 3T and 7T MRI studies^35–39. In contrast, several deep gray matter structures (pallidum, putamen, thalamus, and caudate) exhibited a non-monotonic, U-shaped trajectory, with volumes decreasing from young to middle age (by approximately 2–4%) and partially increasing again in older adults (by 3–6%), resulting in minimal net differences across the lifespan. This pattern aligns with prior 3T MRI reports of subcortical volume variability in the thalamus, pallidum, and striatum across adulthood, reflecting both biological heterogeneity and cohort-related influences associated with motor and cognitive aging^40–42. The ventral diencephalon also decreased by approximately 3%, a magnitude consistent with prior 3T reports of subcortical aging effects^35,36. Cortical regions demonstrated a modest yet consistent reduction, with the cerebral cortex showing a −0.6% decline in older adults relative to the young group. The cerebellar cortex and white matter showed minimal overall change (within  ±1–2%) rather than consistent decline across age groups.

Interestingly, the hippocampus showed minimal overall change (–0.4%), with slightly lower volumes in middle age and a return toward baseline in older adults. While this non-monotonic pattern diverges from the progressive decline typically reported in high-field MRI studies^35,36, it may reflect cohort variability, partial-volume effects, or the resolution limitations inherent to ULF imaging. The amygdala exhibited a similar U-shaped trajectory, with minor midlife reductions (–3.3%) followed by partial recovery in older adults. By contrast, the accumbens area also followed a U-shaped trend, decreasing modestly in middle age (–5.1%) before increasing again in older participants (+0.7%). Overall, these volumetric patterns highlight that ULF MRI-derived templates can capture biologically plausible, though often nonlinear, trajectories of age-related structural variability, such as ventricular enlargement, modest cortical change, and subcortical gray matter heterogeneity, while reflecting the sensitivity limits of 64 mT acquisitions.

Deformation-Based Morphometry Across Age-Group Templates

Deformation-based morphometry (DBM) offers a principled way to examine local morphological change by analyzing the deformation fields required to align anatomical templates. Unlike voxel-based morphometry (VBM), which only measures tissue concentrations after normalization, DBM directly encodes local expansion and contraction via the Jacobian determinants of registration warps¹⁶. DBM has been shown to reliably detect subtle structural changes in both cross-sectional and longitudinal settings without relying on segmentation priors^17,43. In particular, the use of multiple age-specific templates and pairwise DBM is conceptually similar to diffeomorphic aging models, which treat morphology evolution as a smooth deformation trajectory across template space¹⁷. More recently, machine learning-based approaches have incorporated learned aging templates and DBM representations to disentangle normative aging trajectories from disease-related morphological changes⁴⁴.

Hence, to localize spatial patterns of structural change beyond regional volume summaries, we performed DBM between the three age-specific templates. For each pair (Young → Middle, Middle → Old, Young → Old), we registered the older template to the younger template using ANTs SyN (fixed = younger, moving = older), and derived voxel-wise log-Jacobian determinant maps from the forward warp. Under this convention, negative log-Jacobian values (cool colors) indicate that the older template is locally larger and must contract to match the younger (reflecting greater local volume in the older template), whereas positive values (warm colors) indicate that the older template is locally smaller and must expand (reflecting reduced local volume in the older template relative to the younger template).

Figure 6 displays representative sagittal, coronal, and axial slices for T1 (left three columns) and T2 (right three columns) with a shared color scale (log(Jacobian)). As expected, periventricular and CSF spaces show predominantly negative values for the Young ← Middle and Young ← Old registrations, consistent with the age-related ventricular enlargement observed in our volumetric analyses above. Conversely, widespread positive values are observed in cortical and subcortical gray matter, including thalamus, putamen, and pallidum, indicating that older templates are locally smaller, reflecting age-associated tissue contraction, relative to younger. These deformation patterns align with prior 3T MRI findings of ventricular enlargement and subcortical gray matter reduction with aging^35,40,41. The spatial pattern is bilaterally coherent and highly concordant across T1 and T2, supporting the robustness of the ULF-derived templates and the reproducibility of aging effects across modalities.

Fig. 6.

Deformation-based morphometry (DBM) across age-group templates in T1 and T2. Log-Jacobian maps from nonlinear registrations of older → younger templates are shown for three comparisons (rows: Young ← Middle, Middle ← Old, Young ← Old). Columns 1–3 display T1 sagittal, coronal, and axial views; columns 4–6 display the corresponding T2 views. A shared color bar (log(Jacobian)) is used for all panels. Cool colors (negative) indicate regions where the older template must contract to match the younger (older locally larger; e.g., ventricles/CSF), while warm colors (positive) indicate regions where the older template must expand (older locally smaller; e.g., cortex and basal ganglia). The modality-concordant, bilaterally symmetric patterns are consistent with expected ventricular enlargement and gray-matter atrophy with aging.

Data Records

All derived data described in this study are publicly available via Zenodo⁴⁵. The repository provides the population and age-stratified template resources generated from the underlying subject-level cohort. The repository includes:

• Age-specific and population-level group templates.

• Segmented volumes comprising both the original 33 anatomical regions and a merged (left and right hemisphere) set of 18 bilateral regions, where corresponding left and right hemisphere structures were combined to yield single region-wise volumes.

• CSV files containing quantitative volumetric results for all segmented regions across age groups.

All data adhere to FAIR principles, enabling reuse and interoperability with common neuroimaging tools²³. While subject-level datasets provide greater flexibility for statistical and machine learning analyses, template-based resources serve a distinct role by offering standardized spatial references that enable consistent alignment, comparison, and benchmarking across studies. We therefore view subject-level datasets and population templates as complementary resources for the ULF MRI community.

Technical Validation

Several validation steps were performed:

• Bias correction: Improved intensity homogeneity confirmed by visual inspection (Fig. 1).

• Skull stripping: Accurate brain masks across participants, verified against overlaid contours (Fig. 2).

• No post-hoc sharpening: No Laplacian or other edge-enhancing filters were applied; increasing definition across iterations reflects improved nonlinear correspondence and averaging.

• Template construction: Iterative averaging enhanced anatomical sharpness and reduced inter-subject variability (Figs. 3,4).

• Age-group consistency: Age-specific templates demonstrated biologically plausible volumetric patterns, including ventricular enlargement and subtle cortical changes, while subcortical nuclei exhibited non-monotonic (U-shaped) trajectories consistent with midlife volume minima (Table 2, Figs. 5,6). Template-level segmentations were used for qualitative validation and illustrative regional summaries, rather than for subject-level statistical modeling.

• Software-version sensitivity: Data were acquired across multiple scanner software versions; standardized preprocessing reduced large-scale intensity differences, and no systematic effects impacting template construction were observed. To assess whether scanner software version introduced systematic variation in image intensity characteristics, we performed a dedicated software-version sensitivity analysis across the three software versions identified in the DICOM metadata. For both T1- and T2-weighted images, we compared per-image intensity-distribution metrics, including median intensity, interquartile range, and coefficient of variation, at three stages of the preprocessing pipeline: original images, bias-field corrected images, and final skull-stripped, linearly registered images used for template construction. In T1-weighted images, significant software-version-related differences observed in the original and bias-corrected images were no longer significant after the final preprocessing stage, indicating improved comparability after the standardized preprocessing pipeline. In T2-weighted images, preprocessing reduced the magnitude of software-version-related differences, although some residual differences remained for selected intensity metrics. These results suggest that the standardized preprocessing pipeline reduced large-scale software-version-related differences in global intensity characteristics, particularly for T1-weighted images, while some residual differences remained for selected T2-weighted metrics. Full Kruskal–Wallis test results for all modalities, preprocessing stages, and intensity metrics are provided in the GitHub repository associated with this work (see Code Availability).

Together, these evaluations confirm that the resulting templates are anatomically coherent and reproducible, providing a reliable basis for future morphometric and harmonization analyses.

Usage Notes

The templates can be used for:

• Spatial normalization of ULF MRI studies.

• Segmentation and volumetric analyses.

• Training deep learning models for harmonization or enhancement.

• Cross-field adaptation between ULF and high-field MRI.

Users should note the anisotropic acquisition (5 mm slice thickness) inherent to the 64 mT system, which may affect fine-scale morphometry. Resampling mitigates but does not remove this limitation. The templates presented here are derived from data acquired on a single commercially available 64 mT permanent magnet MRI system. Image characteristics in ULF MRI are influenced by scanner-specific factors including magnet geometry, B₀ field homogeneity, gradient nonlinearity, receive coil design, and sequence implementation. As a result, the templates are most directly applicable to datasets acquired using similar hardware and acquisition protocols. We therefore provide the full preprocessing and template construction pipeline to enable re-derivation of population-specific templates for other ULF systems. We further note that scanner-to-scanner variability at ULF may arise from differences in magnet geometry (e.g., Halbach vs other configurations), B₀ field homogeneity, gradient nonlinearity, RF coil geometry, receive bandwidth, patient positioning, and vendor-specific sequence implementation and reconstruction. Despite this specificity, the released templates remain practically useful because the Hyperfine Swoop platform is currently among the most widely deployed ULF systems in human neuroimaging studies, and our pipeline enables re-derivation of templates for other ULF instruments when appropriate.

The applied preprocessing pipeline improves comparability across scans at the cohort level but does not constitute full harmonization of all acquisition- and reconstruction-related variability. A further consideration is that the cohort was acquired across three scanner software versions identified from the DICOM metadata, and the vendor’s proprietary reconstruction pipeline may introduce version-dependent differences in image intensity, noise floor, or apparent sharpness. To evaluate this, we performed a software version sensitivity analysis in the Technical Validation section. Further stratification of the dataset by software version was considered; however, this would substantially reduce the number of subjects within each subgroup and compromise the stability and representativeness of the resulting templates. Instead, we explicitly characterize software-version-related variability and provide the full processing pipeline to enable re-derivation of templates for specific acquisition conditions. This analysis showed that downstream preprocessing substantially reduced software-version-related variability, particularly for T1-weighted images, although some residual differences remained for selected T2-weighted intensity metrics. Accordingly, the templates should be used primarily as ULF-specific spatial references for registration, normalization, and benchmarking under similar acquisition settings, while the accompanying open pipeline enables re-derivation of analogous templates for other software versions, protocols, or ULF systems. Because image reconstruction is vendor-controlled and proprietary, differences between software versions may influence image intensity scaling and noise characteristics, and therefore, quantitative interpretation should be performed with caution, particularly when comparing across independently acquired datasets. We note that recent learning-based approaches, such as SuperSynth, enable super-resolution and registration of low-field images to standard high-field reference spaces using large paired datasets⁴⁶. These methods provide powerful data-driven solutions for cross-field mapping. In contrast, the present work focuses on providing an openly available ULF-specific population template resource and a fully reproducible template-construction pipeline. These contributions are complementary: learning-based methods benefit from standardized reference spaces, while template resources provide interpretable, modality-specific spatial priors independent of model training.

Author contributions

K.T.I. conceived and designed the study, performed data curation, image preprocessing, template construction, statistical analysis, figure generation, and drafted the manuscript. P.Z. and S.Z. contributed to data preprocessing, registration optimization, and quality control. S.D. and H.P. assisted in data management, documentation, and preparation of validation figures. H.K. and P.Z. coordinated data acquisition and clinical protocol standardization at imaging sites. Y.C.F., A.V.D.W., J.F.D.D. and K.C. provided methodological guidance on morphometric analysis and aging interpretation. G.F.E. and M.L. provided oversight, technical review, and critical revision of the manuscript. Z.C. supervised the project, contributed to study design, and provided final manuscript approval. All authors reviewed and approved the final version of the manuscript.

Funding

This research was supported by the National Imaging Facility (NIF), Australia. No specific grant number is associated with this support.

Data Availability

All derived data supporting the findings of this study, including population and age-specific templates, segmentation outputs, and regional volumetric summaries, are openly available on Zenodo at https://doi.org/10.5281/zenodo.17242555.

Code availability

All preprocessing and template construction scripts developed for this study are publicly available at https://github.com/BioMedAnalysis/Template-ULF. The repository includes modular, well-documented Python scripts and shell utilities for each stage of the pipeline, from DICOM conversion to iterative nonlinear template refinement. The implementation relies on the following open-source libraries and software tools:

•dicom2nifti for DICOM-to-NIfTI format conversion (https://pypi.org/project/dicom2nifti/),

• FSL FAST for bias field correction (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FAST),

• FSL BET for skull stripping (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/BET),

• ANTs AffineFast and antsRegistrationSyNQuick[s] for linear and nonlinear registration (https://antspy.readthedocs.io/en/latest/registration.html), and

• SynthSeg⁺ for brain tissue and subcortical segmentation (https://surfer.nmr.mgh.harvard.edu/fswiki/SynthSeg).

Comprehensive usage instructions, example datasets, and configuration templates are provided in the repository to facilitate reproducibility and adaptation of the workflow for other MRI modalities or field strengths.

Competing interests

The authors declare no competing interests.