A comparative study of surrogate musculoskeletal models using various neural network configurations.

摘要

The central idea in musculoskeletal modeling is to be able to predict body-level (e.g. muscle forces) as well as tissue-level information (tissue-level stress, strain, etc.). To develop computationally efficient techniques to analyze such models, surrogate models have been introduced which concurrently predict both body-level and tissue-level information using multi-body and finite-element analysis, respectively. However, this kind of surrogate model is not an optimum solution as it involves the usage of finite element models which are computation intensive and involve complex meshing methods especially during real-time movement simulations. An alternative surrogate modeling method is the use of artificial neural networks in place of finite-element models.;The ultimate objective of this research is to predict tissue-level stresses experienced by the cartilage and ligaments during movement and achieve concurrent simulation of muscle force and tissue stress using various surrogate neural network models, where stresses obtained from finite-element models provide the frame of reference. Over the last decade, neural networks have been successfully implemented in several biomechanical modeling applications. Their adaptive ability to learn from examples, simple implementation techniques, and fast simulation times make neural networks versatile and robust when compared to other techniques. The neural network models are trained with reaction forces from multi-body models and stresses from finite element models obtained at the interested elements. Several configurations of static and dynamic neural networks are modeled, and accuracies close to 93% were achieved, where the correlation coefficient is the chosen measure of goodness. Using neural networks, the simulation time was reduced nearly 40,000 times when compared to the finite-element models. This study also confirms theoretical concepts that special network configurations—including average committee, stacked generalization, and negative correlation learning—provide considerably better results when compared to individual networks themselves.