A Deep Recurrent Neural Network for Predicting Subject-specific Facial Soft Tissue Interaction

Predicting biological soft tissue interaction is of great interest for developing computer-aided decision systems. This study aims to develop and evaluate a novel deep-learning approach based on the recurrent neural network for predicting facial soft tissue impact with a rubber ball. A computational workflow was established including a subject-specific finite element model of the facial soft tissue under interaction with the rubber ball. A series of simulations under different ball velocities was performed to build the learning database. We implemented a long-short-term memory (LSTM) model and then evaluated its performance using root mean square error (RMSE) and regression coefficient metrics. The obtained results showed a RMSE of 3.13 mm and a Pearson correlation coefficient of 0.98 for soft tissue displacement prediction. A RMSE of 0.001 MPa and a Pearson correlation coefficient of 0.94 was also obtained for soft tissue von Mises stress prediction. The present study showed the robustness and accuracy of the recurrent neural network for predicting complex soft tissue interaction behaviours. Our findings open new avenues for deploying novel deep learning workflow for human-facial softtissue interaction. As perspective, this workflow will be integrated into our interactive facial analysis and rehabilitation system.