Physics-Informed Neural Networks for Real-Time Deformation-Aware AR Surgical Tracking

Abstract

Soft tissue deformation severely degrades registration accuracy in AR-assisted surgery. We propose a physics-informed neural network (PINN) that integrates biomechanical priors into real-time depth-based registration. The model embeds finite element elasticity constraints directly into the loss function, allowing neural predictions to remain physically plausible under deformation. Validated on liver and brain phantoms with induced deformations up to 20 mm, the method achieved mean registration error of 1.1 mm, compared with 2.9 mm for conventional ICP and 1.8 mm for FEM-only solvers. Frame rates remained at 22 fps on GPU hardware. Results demonstrate that embedding physics constraints within deep learning significantly enhances robustness in dynamic surgical contexts.