This paper presents a method for simulating the behavior of stomach with large-scale deformation. This simulator is generated by the real-time FEM-based analysis by using a neural network. 4 There are various deformation patterns of hollow organs by changing both its shape and volume. In this case, one network can not learn the stomach deformation with a huge number of its deformation pattern. To overcome the problem, we propose a method of constructing the simulator composed of multiple neural networks by 1)partitioning a training dataset into several subsets, and 2)selecting the data included in each subset. From our experimental results, we can conclude that our method can speed up the training process of a neural network while keeping acceptable accuracy.
INTRODUCITONIn the last decades, there have been developing many Virtual Reality (VR) based systems for supporting surgeons in minimally invasive surgery. The systems use volumetric tissue models to show the virtual images of the inside of a virtual patients' body. Examples of such VR-based systems include a training simulator for endoscopic surgery. 1 The simulator gives surgeons the opportunity of training in endoscopic surgical techniques for various surgery cases. Another application using the model is a navigator for endoscopic surgery. 2 The navigator shows virtual images by merging the models of target tissues into real endoscopic images. The tissue models contain the internal information of the tissues such as the presence of blood vessels and tumors which exist in the tissues. Such information can never be obtained from the endoscopic images. Therefore, the use of the virtual images is a powerful tool for surgeons to approach the tissues or tumors safely and accurately.One of the fundamental components of the support system using the tissue models is to display realistic deformations of tissue models according to their biomechanical behavior. Methods have been proposed for simulating the behavior of soft tissues. 1, 3, 4 Among these methods, the finite element method (FEM) is a well-known technique for accurately modeling the behaviors of continuous objects. Compared with other simulation techniques, FEM can achieve a more physically realistic simulation for deformable objects with linear and nonlinear material properties. On the contrary, the FEM-based simulation of tissue deformation is very time-consuming owing to the need to solve large scale system of equations. Particularly, in the case of soft tissues, the simulation is very complex because the soft tissues have the properties of anisotropy, inhomogeneity, and large-range deformation.To solve these problems, Morooka et al. 4 proposed a real-time FEM-based simulator for deforming a soft tissue model by using a neural network. To construct the simulator, the training dataset is generated by collecting many training data. Each data is a pair of an external force caused by the contact between virtual surgical instruments and tissue models and