Dual-branch network-based simulation of time-series deformation for liver

Because of liver viscoelasticity, its stress-strain response depends not only on the direction of the applied force but also on the duration of the force. When the external force is first applied, the liver will show a large deformation rate, and the strain will gradually stabilize with the increase of time, which increases the complexity of the simulation. We propose a recursion-based dual-branch network that can effectively deal with the feature differences at different stages of time series and accurately predict the deformation of the liver under continuous external forces. We adopted a scheduled sampling strategy to alleviate the exposure bias caused by training the model only with gold standard data. In addition, we propose an incremental-global loss function that can capture subtle changes at the current moment while maintaining the stability of long-term predictions. The training set is constructed by applying external forces in random directions to 30 randomly selected points on the surface of the liver. For validation, we selected an additional 10 points and applied random external forces in the same pattern. The experimental results show that our method has higher prediction accuracy than three commonly used time series prediction models.