New MRI-to-Model Pipeline Builds Personalized Brain Simulations from Rugby Player Scans

Researchers in New Zealand have built a fully reproducible computational pipeline that converts magnetic resonance imaging (MRI) scans of adolescent rugby players into detailed finite element (FE) head models, offering a new way to study how the brain deforms during head impacts. The framework, published in Annals of Biomedical Engineering, addresses a persistent problem in traumatic brain injury (TBI) research: most existing FE head models are derived from a single subject’s scan, a generic 50th-percentile brain, or a broad open-source atlas that may poorly represent the specific population being studied. By generating a cohort-specific brain template from 78 adolescent male collision sport athletes and then converting it into a biomechanically detailed FE model, the team has demonstrated that both the anatomy and the microstructure of the brain can be carried directly from imaging data into simulation.

The work was led by Christian John A. Saludar, Maryam Tayebi, Eryn Kwon, and colleagues, including the Mātai mTBI Research Group, as part of an ongoing longitudinal study of adolescent male rugby union athletes conducted in New Zealand under ethical approval from the New Zealand Health and Disability Ethics Committee. The researchers recruited participants across multiple competitive secondary school seasons, excluding those with pre-existing neuropsychiatric conditions, dental braces that could cause imaging artifacts, or a self-reported mild TBI within six months of the study season. Each athlete underwent multimodal MRI scanning at early-, mid-, and post-season time points using a 3.0 Tesla scanner equipped with a 48-channel head coil. For the template construction, only early-season scans were included, a deliberate choice to control for any microstructural alterations that repetitive subclinical head acceleration events might induce throughout a playing season.

The imaging protocol combined high-resolution T1-weighted structural scans with multi-shell diffusion-weighted MRI, capturing both brain geometry and white matter microstructure. After extensive pre-processing to remove noise, correct field inhomogeneities, and address motion and eddy-current artifacts, the team generated what they call the AMARA-78 template — short for ABI-Matai Adolescent Rugby Atlas 78 — by iteratively averaging and warping each participant’s T1-weighted and fractional anisotropy (FA) images into a common reference space. The process used the ANTs software package with a Greedy SyN transformation model, continuing until the Pearson cross-correlation between successive template iterations reached a plateau. A separate population fiber orientation distribution (FOD) template was also generated using MRtrix3, capturing the average white matter fiber architecture across the cohort.

To evaluate whether this new template actually represents the study population better than existing alternatives, the researchers performed deformation-based morphometry on a separate group of 20 male non-collision sport athletes who served as controls. These participants played sports including basketball, track and field, rowing, and waka ama (outrigger canoe racing). When the control brains were registered to the AMARA-78 template, fewer voxels showed significantly large local deformation compared with registration to the widely used MNI152 atlas (94,920 voxels for AMARA-78 versus 140,148 for MNI152) or to the Purdue Neurotrauma Group (PNG) adolescent atlas (169,762 versus 376,624 voxels). The MNI152 template was built from adults aged 18 to 44 years, while the PNG atlas, though derived from adolescent collision sport athletes, still produced greater deformation than AMARA-78. The authors suggest that factors beyond age, including demographic differences, contribute to morphological variability across cohorts.

Once the template was established, the team converted it into a finite element head model using a series of open-access tools. A surface mesh was generated from the template brain mask using 3DSlicer, then imported into PolyCube to create a hexahedral volume mesh compatible with the open-source FE solver FEBio. The final model contains 92,736 elements with edge lengths ranging from 1.5 to 5.0 millimeters. An in-house Python script established direct correspondence between FE elements and MRI voxels, allowing tissue class assignments — gray matter, white matter, and cerebrospinal fluid — and spatially varying material properties to be mapped element by element from the imaging data. White matter was modeled using a hyper-viscoelastic fiber-reinforced anisotropic formulation based on the Gasser-Ogden-Holzapfel (GOH) model, with the fiber dispersion parameter informed by FA values and viscoelastic relaxation parameters drawn from prior literature. The falx cerebri and tentorium cerebelli were represented as manually defined membrane structures between the cerebral hemispheres and between the cerebrum and cerebellum.

Validation against cadaveric head impact experiments yielded small to moderate CORA scores, a standard metric for comparing simulated and experimental biomechanical responses. Against the pure coronal rotation experiments of Alshareef et al., which applied rotational loading to a 53-year-old male cadaveric head at a peak angular velocity of 20 radians per second, CORA scores ranged from 0.26 to 0.56. Against the six-degree-of-freedom experiments of Hardy et al., which used an 82-year-old male specimen, scores ranged from 0.46 to 0.82. The authors note that these scores are comparable to those reported for established FE head models in the literature, and that the qualitative displacement magnitudes — approximately 0 to 3 millimeters — closely matched the experimental data. They also acknowledge that the age and geometric differences between the adolescent model and the older cadaveric specimens likely contribute to the deviation, and that specimen- or population-matched experimental data would be needed to fully assess the benefit of personalization.

A sensitivity analysis varying input material properties by ±10 percent revealed that the densities of gray matter and white matter exerted the greatest influence on validation performance against the Alshareef et al. experiment, followed by the Poisson’s ratio of the dura, falx, and tentorium. The researchers also compared the FE model volume against MRI-derived total intracranial volume across the rugby cohort, finding a strong correlation (r = 0.998, p < 0.05), with the AMARA-78 model volume at 1,574,149 cubic millimeters and the mean subject FE model volume at 1,592,028.81 ± 107,265.25 cubic millimeters.

To test whether the pipeline could generate subject-specific models, the researchers randomly selected one athlete from the cohort and morphed the template FE mesh into that individual’s brain geometry using a free-form deformation algorithm. The morphing preserved the nodal and elemental relationships of the template mesh, enabling direct inter-subject comparison in future large-cohort studies. The team then applied two loading conditions derived from instrumented mouthguard data pooled across the entire season: the 95th percentile head acceleration event, with a peak angular acceleration of 3,416 radians per second squared, and the 99th percentile event, reaching 6,371 radians per second squared. Three model configurations were compared: an isotropic template model, a template model incorporating FA-derived fiber dispersion parameters, and the fully subject-specific model.

The results showed that the subject-specific model produced significantly different regional maximum principal strain distributions compared with both template-based models under both impact intensities (p < 0.01, after false discovery rate correction). In contrast, the two template models — isotropic and anisotropic — did not differ significantly from each other, although the anisotropic formulation predicted higher strain magnitudes, particularly at later time points. Regional strains across all models generally peaked between 20 and 30 milliseconds before decreasing at 40 milliseconds. A linear mixed-effects model accounting for the nested structure of 84 brain regions across four time points confirmed that model type and timestep both significantly influenced predicted strain. The authors suggest that the difference between the subject-specific and template models may be driven more by intracranial volume differences than by variation in material properties, consistent with prior findings that brain size influences mechanical response.

The study has several limitations. The cohort was exclusively male, limiting the ability to assess sex-specific differences in brain response, though the pipeline itself is designed to be applicable to cohorts defined by age, sex, ethnicity, or other population characteristics. The AMARA-78 template showed deviations from the cohort mean in cerebrospinal fluid and bilateral cerebral cortex volumes, attributed to nonlinear registration and intensity averaging across subjects with varying sulcal geometry. No additional event-level verification or filtering of the instrumented mouthguard data was performed beyond the manufacturers’ processing pipelines. The subject-specific comparison was limited to a single athlete, and the authors emphasize that further investigation into the added value of subject-specific FE models is warranted. Despite these constraints, the framework’s reliance on open-access packages across the full workflow — from MRI pre-processing through FE simulation — means that other research groups can replicate and adapt it to their own cohorts without being limited by access to proprietary algorithms, potentially addressing the shortage of biomechanically representative models across different sports, sexes, and age groups in TBI research.

Subject of Research: A reproducible MRI-to-finite-element workflow for generating cohort-specific and subject-specific human head models from adolescent rugby player brain imaging

Article Title: A Reproducible MRI-to-FE Workflow for Generating Population-Specific and Subject-Specific Finite Element Head Models

Article References: Saludar, C. J. A., Tayebi, M., Kwon, E., Mathew, J., McGeown, J., Schierding, W., Mātai mTBI Research Group, Potter, L., Condron, P., Taylor, D., Cornfield, D., McHugh, P., Emsden, T., Danesh-Meyer, H., MacDonald, M., Newburn, G., Nepe-Apatu, T., Bydder, G., Wang, A., … Shim, V. (2026). A Reproducible MRI-to-FE Workflow for Generating Population-Specific and Subject-Specific Finite Element Head Models. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04432-y

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04432-y

Keywords: finite element modeling, traumatic brain injury, mild TBI, MRI, diffusion-weighted imaging, brain atlas, rugby, head acceleration events, biomechanics, subject-specific modeling, white matter anisotropy, instrumented mouthguard