[Paper Review] Towards Fast Biomechanical Modeling of Soft Tissue Using Neural Networks

Felix Meister, Tiziano Passerini+4 more|arXiv (Cornell University)|Dec 13, 2018

Elasticity and Material Modeling10 references4 citations

TL;DR

This paper proposes a deep neural network to directly predict vertex-wise accelerations in soft tissue deformations over large time steps, bypassing traditional explicit finite element method (FEM) stability limits. Trained on a single synthetic bar deformation under torsion, the model achieves sub-millimeter accuracy (0.017 mm ± 0.014) at 10× the original time step and remains stable beyond the FEM stability limit, though artificial stiffening emerges at very high time steps.

ABSTRACT

To date, the simulation of organ deformations for applications like therapy planning or image-guided interventions is calculated by solving the elastodynamics equations. While efficient solvers have been proposed for fast simulations, methods that are both real-time and accurate are still an open challenge. An ideal, interactive solver would be able to provide physically and numerically accurate results at high frame rate, which requires efficient force computation and time integration. Towards this goal, we explore in this paper for the first time the use of neural networks to directly learn the underlying biomechanics. Given a 3D mesh of a soft tissue segmented from medical images, we train a neural network to predict vertex-wise accelerations for a large time step based on the current state of the system. The model is trained using the deformation of a bar under torsion, and evaluated on different motions, geometries, and hyperelastic material models. For predictions of ten times the original time step we observed a mean error of 0.017mm $\pm$ 0.014 (0.032) at a mesh size of 50mm x 50mm x 100mm. Predictions at 20dt yield an error of 2.10mm $\pm$ 1.73 (4.37) and by further increasing the prediction time step the maximum error rises to 38.3mm due to an artificial stiffening. In all experiments our proposed method stayed stable, while the reference solver fails to converge. Our experiments suggest that it is possible to directly learn the mechanical simulation and open further investigations for the direct application of machine learning to speed-up biophysics solvers.

Motivation & Objective

• Address the challenge of real-time, accurate soft tissue deformation simulation for medical applications like therapy planning and image-guided interventions.
• Overcome the trade-off between computational speed and accuracy in traditional biomechanical solvers, particularly explicit FEM which requires small time steps for stability.
• Explore whether deep learning can directly approximate the underlying biomechanical dynamics, enabling large time-step predictions.
• Develop a data-driven surrogate model that generalizes across different geometries, motions, and hyperelastic material models.
• Demonstrate stability and accuracy beyond the stability limit of explicit FEM using a learned acceleration predictor.

Proposed method

• Train a fully connected feedforward neural network to predict vertex accelerations for a large time step Δt based on the current system state: displacement u(t), velocity u̇(t), and total force f_T(t) = f_e - Ku.
• Transform input features into a local coordinate system to improve generalization and invariance to global motion.
• Use a synthetic dataset generated via total Lagrangian explicit dynamics (TLED) FEM with a hyperelastic exponential material model (Ψ = a/(2b)exp(b(I₁−3)) + d(J−1)).
• Train the network end-to-end using regression loss between predicted and ground-truth accelerations from TLED simulations.
• Apply the trained network to simulate deformations at Δt = 10dt, 20dt, 50dt, and 100dt, with a 'burn-in' phase using TLED to initialize the state.
• Update displacement using a second-order time integration scheme: u(t+Δt) = ü_Δt(t)·Δt² + 2u(t) - u(t−Δt).

Experimental results

Research questions

• RQ1 Can a deep neural network be trained to directly predict large-time-step accelerations in soft tissue deformation, bypassing explicit FEM time-step constraints?
• RQ2 How well does the learned model generalize to unseen geometries (e.g., cylinder, liver lobe) and motion types (torsion, stretch, compression) beyond the training data?
• RQ3 What is the trade-off between prediction accuracy and time-step size when the model exceeds the stability limit of explicit FEM?
• RQ4 Does the model exhibit stable behavior at extreme time steps (e.g., 50dt, 100dt) where traditional FEM fails to converge?
• RQ5 To what extent does the model capture complex material behavior, such as Neo-Hookean response, without retraining?

Key findings

• At 10× the original time step (10dt), the model achieved a mean prediction error of 0.017 mm ± 0.014 (0.032) compared to the FEM ground truth, demonstrating high accuracy.
• At 20dt, the error increased to 2.10 mm ± 1.73 (4.37), indicating the onset of artificial stiffening, though the simulation remained stable.
• For 50dt and 100dt predictions, the model remained numerically stable and did not diverge, unlike the reference FEM solver which would fail to converge.
• The maximum error reached 38.3 mm at 100dt, primarily due to cumulative artificial stiffening, suggesting limitations in network capacity and training data diversity.
• The model generalized well across different motions (torsion, stretch, compression), geometries (bar, cylinder, liver lobe), and material models (exponential, Neo-Hookean).
• The results suggest that the proposed AI-based solver combines the parallelizability of explicit FEM with the stability of implicit FEM, enabling large time-step simulations.