Shahbazi, Ashkan; Pereira, Kyvia; Heiselman, Jon S.; Akbari, Elaheh; Benson, Annie C.; Seifi, Sepehr; Liu, Xinyuan; Johnston, Garrison Lawrence Horswill; Wu, Jie Ying; Simaan, Nabil; Miga, Michael; Kolouri, Soheil (2026).听.听Proceedings of the AAAI Conference on Artificial Intelligence, 40(30), 25243鈥�25251.
Accurately simulating how聽soft tissues聽(like organs) deform during surgery is important for applications such as surgical training, robotic surgery, and automated surgical systems. Traditional physics-based methods like the聽Finite Element Method (FEM)聽can model these deformations well, but they are often too slow for real-time use. Faster,聽neural network鈥揵ased models聽can make quick predictions, but if they rely only on data and ignore physical laws, they may produce unrealistic results.
This study introduces a new hybrid approach that combines the speed of neural networks with built-in physics knowledge. Specifically, it incorporates聽Kelvinlets聽(mathematical solutions that describe how materials deform under force) as guiding principles, along with large datasets generated from FEM simulations. This allows the model to capture both simple and complex tissue behaviors while staying physically realistic.
The method can quickly and accurately predict how tissues deform when manipulated by surgical tools, including cases where multiple tools interact at once. Tests on simulated surgical tasks show that this approach produces more stable and realistic results than existing methods, while still running fast enough for real-time applications. Overall, the work demonstrates a promising way to build fast, reliable simulations for use in surgical AI and robotics.
