PerSiVal: deep neural networks for pervasive simulation of an activation-driven continuum-mechanical upper limb model

Area of Science:

• Biomechanics
• Computer Science
• Medical Simulation

Background:

• Continuum-mechanical musculoskeletal models offer detailed simulations but require significant computational resources.
• Real-time visualization on resource-poor systems (AR, mobile) is crucial for human-machine interaction, surgical support, and physiotherapy.

Purpose of the Study:

• To develop a novel deep neural network architecture for real-time visualization of musculoskeletal system simulations.
• To enable the deployment of complex biomechanical models on resource-limited platforms.

Main Methods:

• A densely connected neural network was designed for visualizing continuum-mechanical musculoskeletal simulations.
• An activation-driven five-muscle upper limb model was used to obtain muscle deformations.
• A sparse grid surrogate captured surface deformation, and a deep learning model was trained for real-time visualization.

Main Results:

• The deep neural network achieved an average positional error of 0.97±0.16 mm (0.57±0.10%) for the biceps brachii surface mesh.
• Evaluation times were 9.88 ms (CPU) and 3.48 ms (GPU), enabling theoretical frame rates of 101 fps and 287 fps.
• The method proved effective for real-time visualization of complex meshed models.

Conclusions:

• The combination of surrogates and deep neural networks enables real-time visualization of complex musculoskeletal models.
• This approach is applicable to other real-time visualizations of complex meshed models beyond the musculoskeletal system.