Longitudinal Biomechanical Reference States

Updated 6 April 2026

Longitudinal Biomechanical Reference States are detailed, time-indexed models that characterize tissue mechanics and morphology using finite-deformation and nonlinear elasticity.
They integrate constitutive hyperelastic laws, data-driven growth maps, and iterative computational algorithms to simulate tissue evolution.
Applications include predictive modeling, counterfactual analysis, and early detection of pathologies in personalized clinical and research settings.

Longitudinal biomechanical reference states constitute a rigorous, time-indexed characterization of biological tissue mechanics and morphology grounded in explicit biomechanical models. They capture normative, perturbative, or pathological evolution of structure and function at individual or population levels. Such reference states arise through the integration of constitutive hyperelastic laws, subject/disease-specific growth or atrophy maps, and empirical or imaging-based calibration. This approach enables quantitative comparison, prediction, and counterfactual modeling of tissue deformation trajectories.

1. Mathematical Formulation and Biomechanical Principles

The foundation of constructing longitudinal biomechanical reference states is the formalism of finite-deformation, nonlinear elasticity. The total deformation gradient $F$ is multiplicatively decomposed into a growth (or atrophy) component $F_g$ and an elastic correction $F_e$ : $F = F_e \cdot F_g$ $F_g$ prescribes tissue-specific volumetric or anisotropic expansion/shrinkage derived from longitudinal data (e.g., growth laws, atrophy regressors), while $F_e$ enforces mechanical equilibrium under an appropriate hyperelastic energy density. For isotropic processes and $d$ spatial dimensions,

$F_g(x) = a(x)^{-1/d}I$

where $a(x)$ is the local volume-ratio field.

The equilibrium equations arise from minimizing the total free energy, typically manifesting as the balance of the first Piola–Kirchhoff stress: $\text{div}\,P = 0$ Under a Neo-Hookean constitutive law, the strain energy $F_g$ 0 and stress $F_g$ 1 have the form: $F_g$ 2 where $F_g$ 3, $F_g$ 4 is the shear modulus, and $F_g$ 5 is the bulk modulus with $F_g$ 6 in brain tissue modeling (Silva et al., 2021, Baena et al., 13 Aug 2025, Silva et al., 2020, Wang et al., 2019).

2. Reference State Construction: Frameworks and Algorithms

Reference states are operationalized differently across biological applications but universally rely on explicit model parameterization and iterative temporal updating.

Brain tissue atrophy (Da Silva et al.): A multilayer perceptron (MLP) regresses the regional atrophy field $F_g$ 7 from demographic covariates $F_g$ 8, assigning $F_g$ 9 to each brain parcel. This atrophy informs a spatial field $F_e$ 0 and drives the hyperelastic solver to compute updated tissue positions $F_e$ 1, generating time-indexed volume and strain curves $F_e$ 2 (Silva et al., 2021).
Neurodevelopmental modeling (NEUBORN): Twin hierarchical U-Nets predict stationary velocity fields at multi-scale, enforcing biomechanical regularization via volumetric Neo-Hookean energy. Diffeomorphic integration of velocities enables interpolation of deformations across arbitrary intermediate times, yielding individualized, biologically-plausible growth trajectories and reference states $F_e$ 3 for any $F_e$ 4 (Baena et al., 13 Aug 2025).
Biomechanical deep registration (Barral et al.): A U-Net architecture learns to map prescribed atrophy/growth fields to displacement fields, enabling rapid generation of continuous families of plausible deformed states $F_e$ 5. Training loss merges hyperelastic energy, boundary constraints, and optionally image fidelity to ensure mechanical realism (Silva et al., 2020).
Explicit FE growth steps (Wang et al.): At each discrete time increment, the stress-free configuration is updated by applying $F_e$ 6 based on empirical growth laws (e.g., quadratic fits to brain longitudinal length), followed by stress relaxation via elastic equilibrium, and then re-declaration of the new configuration as stress-free. This “reset” paradigm avoids residual stress accumulation and ensures parameter-fit to in vivo metrics (Wang et al., 2019).

3. Data-Driven Inputs and Regression of Growth/Atrophy Maps

Key to constructing biologically informative reference states is the derivation of $F_e$ 7 from empirical or data-driven models.

Demographic Regressors: In brain atrophy simulation, MLPs map demographic variables to region-wise atrophy factors $F_e$ 8, supporting continuous-time predictions for both healthy and pathological states. For instance, hippocampal engineering strain curves for healthy controls and Alzheimer's disease are given as

$F_e$ 9

and analogous regional parametrizations exist for cortex and ventricles (Silva et al., 2021).

Image-Based Growth Parameterization: In NEUBORN, isotropic growth $F = F_e \cdot F_g$ 0 is computed from segmentations at serial timepoints, forming the prescribed $F = F_e \cdot F_g$ 1 for each tissue. The registration loss penalizes deviation from both data fidelity and biomechanical feasibility (Baena et al., 13 Aug 2025).
Longitudinal Physical Laws: For developmental models, anatomical metrics such as brain longitudinal length (BLL) or cortical thickness are fit to normative datasets (e.g., BLL $F = F_e \cdot F_g$ 2 in gestational weeks) and used to drive isotropic or tangential growth in finite-element models (Wang et al., 2019).

4. Validation Metrics and Constraints for Biomechanical Plausibility

Reference states must be validated both biomechanically and anatomically:

Topological Preservation: Negative Jacobian determinants ( $F = F_e \cdot F_g$ 3) are non-physical (folding/tearing) and are minimized via appropriate loss terms. For example, NEUBORN reduces negative-Jacobian voxels by ~ $F = F_e \cdot F_g$ 4 versus unconstrained baselines (Baena et al., 13 Aug 2025).
Dice Coefficient and ASPVC: To assess anatomical overlap and volume change, metrics such as Dice similarity coefficient (87.3 ± 0.16 % for NEUBORN), and Absolute Symmetric Percentage Volume Change (e.g., NEUBORN 3.82 ± 0.24 %) are employed. These quantify the biological plausibility and adherance to population-level growth trends (Baena et al., 13 Aug 2025).
Quantitative Error Metrics: Atrophy map MSE (5.1×10⁻⁵), image MSE, and regional Dice for tissue types are used to confirm that biomechanical models match empirical ground truth with high fidelity (Silva et al., 2020).
Regional and Bilateral Asymmetry: For in vivo measurement (e.g., corneal biomechanics), regional differences in modulus and shifts of >20 MHz, and bilateral asymmetry >12 MHz (cone region) serve as robust indicators for pathological changes (e.g., keratoconus) (Shao et al., 2018).

5. Applications and Interpretative Use of Reference States

Longitudinal biomechanical reference states enable a spectrum of applications:

Normative Trajectories: By computing reference strain or volume curves (e.g., $F = F_e \cdot F_g$ 5, $F = F_e \cdot F_g$ 6), expected biomechanical evolution can be defined for healthy versus disease groups, supporting early deviation detection and risk stratification (Silva et al., 2021, Baena et al., 13 Aug 2025).
Counterfactual and Interventional Modeling: By varying inputs such as disease class or synthetic atrophy rates, counterfactual deformation fields simulate hypothetical interventions (e.g., slowing AD progression by reducing $F = F_e \cdot F_g$ 7) (Silva et al., 2021).
Predictive Forward Modeling: Extrapolation of learned deformation fields beyond observed intervals allows prediction of future tissue configuration, relevant in forecasting neurodevelopmental trajectories or anticipating malformations (Baena et al., 13 Aug 2025).
Personalized Baselines and Early Diagnosis: Reference states establish individualized and population-level mechanical baselines (e.g., corneal modulus maps, brain volumetric trends). Deviations from these baselines can be flagged as early biomarkers (e.g., focal cone-region weakening in keratoconus, rate shifts in AD) (Shao et al., 2018, Silva et al., 2021).

6. Biological Validation and Calibration

Accurate reference states depend critically on empirical calibration:

Clinical Datasets: Validation against large-scale, longitudinal MRI data (ADNI, dHCP) ensures that simulated structural evolutions mirror observed population biology (Silva et al., 2021, Baena et al., 13 Aug 2025, Silva et al., 2020).
Numerical Fitting: Growth laws and atrophy regressors are fit to longitudinal series, with high $F = F_e \cdot F_g$ 8 (e.g., BLL curve $F = F_e \cdot F_g$ 9 for fetal brain length) confirming the empirical adequacy (Wang et al., 2019).
Mechanical Parameterization: Material properties (e.g., $F_g$ 0 and $F_g$ 1 values, modulus ranges from Brillouin shifts) are set and validated by direct measurement or literature consensus (e.g., $F_g$ 2 in GM/WM, $F_g$ 3 in CSF for brain; $F_g$ 4 GPa in healthy cornea) (Silva et al., 2021, Shao et al., 2018).

7. Extensions, Generalizations, and Limitations

The reference state framework generalizes to any soft-tissue growth or atrophy scenario with explicit biomechanical and empirical constraints:

Other Organs and Pathologies: The same principles apply to cardiac growth, tumor morphogenesis, tissue engineering constructs, and plant organogenesis by selecting appropriate $F_g$ 5, time-dependent morphological metrics, and boundary conditions (Wang et al., 2019).
Modeling Pathological or Interventional Change: Disease-related parameter deviations or therapeutic interventions are readily modeled, with resultant deformation histories compared against reference trajectories (Silva et al., 2021, Shao et al., 2018).
Limitations: Biomechanical reference states depend on the accuracy of constitutive models, empirical fits, and the biological validity of imposed constraints. Non-physical behaviors (e.g., folding) remain a technical challenge, mitigated by topology-preserving regularization (Baena et al., 13 Aug 2025). Biological phenomena outside the mechanical model's expressiveness (e.g., biochemical drivers, remodeling, poroelasticity) may not be captured unless explicitly included.