Personalization of human body models and beyond via image registration

📝 Paper Summary

Biomechanics simulation Finite Element Human Body Models (HBM)

A landmark-free method personalizes finite element human body models by converting them into binary images and using Diffeomorphic Demons registration to compute deformation fields for mesh morphing.

Core Problem

Generating personalized Human Body Models (HBMs) typically relies on landmark-based methods (RBF, Kriging) which are time-consuming, require manual correspondence of anatomical points, and are computationally expensive for dense surfaces.

Why it matters:

Personalized models are essential for reconstructing accidents and understanding injury mechanisms in diverse populations (age, sex, BMI)
Developing validated HBMs from scratch is prohibitively expensive; efficient morphing of baseline models is required to scale simulations
Manual landmarking in existing methods limits automation and scalability when processing large datasets of subjects

Concrete Example: When morphing a baseline 50th-percentile male model to a high-BMI subject, traditional parametric methods lacking skeleton data fail to capture the specific subcutaneous fat distribution, leading to unrealistic expansions. This method uses voxelized skeleton and skin targets to accurately align internal structures.

Key Novelty

Image Registration-Based Mesh Morphing

Converts 3D mesh models (skin and skeleton) into binary voxel images, treating the geometry alignment problem as an image registration task
Uses the Diffeomorphic Demons algorithm to calculate a dense displacement field for every voxel, which is then interpolated to move the Finite Element nodes to their new positions

Evaluation Highlights

Achieved a mean DICE score of 0.94 across 10 personalized models (morphing baselines to subjects with varying BMI and sex)
VIVA+ female-to-male morphing achieved 0.96 DICE and 5.7 mm HD95 (Hausdorff Distance), comparable to RBF-based baselines
Maintained element quality (Jacobian, aspect ratio) comparable to baseline models, enabling direct runnability in 40 km/h side impact simulations

Breakthrough Assessment

7/10

Offers a significant workflow improvement (landmark-free, automated) for biomechanics researchers. While the underlying registration algorithm (Demons) is established, applying it to whole-body FE morphing is a novel and practical contribution.

⚙️ Technical Details

Problem Definition

Setting: Deform a baseline Finite Element mesh $M_{base}$ to match a target subject geometry $S_{target}$ while preserving mesh topology and element quality.

Inputs: Baseline HBM surface (skin + skeleton), Target subject surface (skin + skeleton)

Outputs: Personalized HBM mesh with updated nodal coordinates

Pipeline Flow

Pre-processing: Voxelize HBM and Subject -> Binary Images
Registration: Demons Registration -> Displacement Field
Morphing: Interpolate Field -> Update Mesh Nodes

System Modules

Voxelizer

Convert surface meshes (STL/OBJ) of skin and skeleton into binary 3D images

Model or implementation: Slicer 3D 'Convert model to segmentation node'

Demons Registrar

Compute non-linear deformation field to align baseline image to target image

Model or implementation: Diffeomorphic Demons Algorithm (implemented in Slicer 3D)

Mesh Morpher

Apply displacement field to Finite Element nodes to generate personalized model

Model or implementation: Linear Interpolation

Novel Architectural Elements

Application of voxel-based Diffeomorphic Demons registration to FE mesh morphing (typically a geometric interpolation task)
'Shielding' pipeline (Type III) which replaces parts of the target image with the baseline's own geometry to prevent morphing in specific regions (e.g., head)

Modeling

Base Model: Evaluated on SAFER HBM, THUMS v4.02, VIVA+ (Seated/Standing)

Key Hyperparameters:

smooth_factor: 2

Compute: Fast execution (minutes for registration/morphing modules); Pre-processing requires manual conversion steps.

Comparison to Prior Work

vs. RBF/Kriging: Landmark-free (saves manual effort) and handles dense volumetric changes naturally via voxelization
vs. PIPER: Does not require metadata files with pre-defined landmarks
vs. SSM: Can personalize to specific individual scans/shapes rather than just statistical parameters [not cited in paper as direct baseline, but conceptual difference]

Limitations

Registration accuracy depends on the overlap of source and target images; limbs must be pre-positioned to similar postures
Skeleton accuracy (e.g., ribcage) in the personalized model can deviate if the registration regularization (smooth factor) is too high
Introduction of intersecting contact surfaces in some cases (e.g., 8 intersections for SAFER to subj1), requiring manual repair
Limited capability to morph distinct internal organ shapes if they differ significantly between baseline and subject

Reproducibility

Method relies on open-source software (3D Slicer) and public HBMs (VIVA+, THUMS). Anthropometric data for subjects provided. Specific scripts to automate the Slicer-to-FE pipeline are not explicitly linked but described as available from authors.

📊 Experiments & Results

Evaluation Setup

Morphing 4 baseline HBMs into 10 target subjects (varying height 160-187cm, BMI up to 34, Sex)

Benchmarks:

Personalization Accuracy (Geometric Alignment) [New]
Element Quality (Mesh Validity Check)
Runnability (FE Simulation Stability)

Metrics:

DICE coefficient (Overlap)
HD95 (Hausdorff Distance 95th percentile)
Jacobian (Element Quality)
Aspect Ratio (Element Quality)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
10 Subject Set	DICE (Mean)	Not reported in the paper	0.94	Not reported in the paper
10 Subject Set	HD95 (Mean)	Not reported in the paper	10.63	Not reported in the paper
VIVA+ Female to Male (Standing)	DICE	0.97	0.96	-0.01
VIVA+ Female to Male (Standing)	HD95	Not reported in the paper	5.72	Not reported in the paper

Main Takeaways

The method successfully morphs diverse baselines (SAFER, THUMS, VIVA+) to diverse subjects (BMI 34, short stature) with high geometric fidelity (DICE > 0.9)
Element quality (Jacobian, Aspect Ratio) is preserved relative to the baseline, enabling simulations without extensive remeshing
Geometric correction capability allows 'shielding' specific parts (like the head) to correct artifacts or anatomical anomalies in baselines
Skeleton alignment is generally accurate (avg error ~2.4mm for VIVA+ case) but can show larger deviations in ribs/hips compared to skin

📚 Prerequisite Knowledge

Prerequisites

Finite Element Method (FEM) basics
Image Registration principles (fixed vs. moving images)
Medical image processing (voxelization)

Key Terms

HBM: Human Body Model—a computational finite element model representing human anatomy for crash safety simulations

FE: Finite Element—a numerical method for solving physical problems by subdividing a large system into smaller, simpler parts (elements)

Image Registration: The process of transforming different sets of data into one coordinate system; here, aligning the baseline model image to the target subject image

Demons Algorithm: A non-linear image registration technique that computes displacement fields by treating image intensity differences as forces pushing the 'demons' (pixels) to align boundaries

Voxelization: Converting a geometric object (like a 3D mesh) into a 3D grid of values (voxels), similar to pixelating a 2D image

DICE: Sørensen–Dice coefficient—a statistic used to gauge the similarity of two samples (overlap between volumes), ranging from 0 (no overlap) to 1 (perfect overlap)

HD95: 95th-percentile Hausdorff Distance—a metric measuring the maximum distance between two contours, robust to outliers by taking the 95th percentile

Jacobian: In FE analysis, a measure of element distortion; a Jacobian of 1 indicates a perfect element, while values <= 0 indicate inversion/failure

RBF: Radial Basis Function—a mathematical interpolation method used in traditional morphing to deform meshes based on sparse landmark points

Kriging: A geostatistical interpolation method used in the PIPER software for mesh deformation based on control points