Efficient techniques for soft tissue modeling and simulation
Performing realistic deformation simulations in real time is a challenging problem in computer graphics. Among numerous proposed methods including Finite Element Modeling and ChainMail, we have implemented a mass spring system because of its acceptable accuracy and speed. Mass spring systems have, however, some drawbacks such as, the determination of simulation coefficients with their iterative nature. Given the correct parameters, mass spring systems can accurately simulate tissue deformations but choosing parameters that capture nonlinear deformation behavior is extremely difficult. Since most of the applications require a large number of elements i. e. points and springs in the modeling process it is extremely difficult to reach realtime performance with an iterative method. We have developed a new parameter identification method based on neural networks. The structure of the mass spring system is modified and neural networks are integrated into this structure. The input space consists of changes in spring lengths and velocities while a "teacher" signal is chosen as the total spring force, which is expressed in terms of positional changes and applied external forces. Neural networks are trained to learn nonlinear tissue characteristics represented by spring stiffness and damping in the mass spring algorithm. The learning algorithm is further enhanced by an adaptive learning rate, developed particularly for mass spring systems. In order to avoid the iterative approach in deformation simulations we have developed a new deformation algorithm. This algorithm defines the relationships between points and springs and specifies a set of rules on spring movements and deformations. These rules result in a deformation surface, which is called the search space. The deformation algorithm then finds the deformed points and springs in the search space with the help of the defined rules. The algorithm also sets rules on each element i. e. triangle or tetrahedron so that they do not pass through each other. The new algorithm is considerably faster than the original mass spring systems algorithm and provides an opportunity for various deformation applications. We have used mass spring systems and the developed method in the simulation of craniofacial surgery. For this purpose, a patient-specific head model was generated from MRI medical data by applying medical image processing tools such as, filtering, the segmentation and polygonal representation of such model is obtained using a surface generation algorithm. Prism volume elements are generated between the skin and bone surfaces so that different tissue layers are included to the head model. Both methods produce plausible results verified by surgeons.