In virtual surgical training, it is crucial to achieve real-time, high-fidelity simulation of the tissue deformation. The anisotropic and nonlinear characteristics of the organ with multi-component make accurate real-time deformation simulation difficult. A localized tissue constraint progressive transfer learning method is proposed in this paper, where the base-compensated dual-output transfer learning strategy and the localized tissue constraint progressive learning architecture are developed. The proposed strategy enriches the multi-component biomechanical dataset to fully represent complex force-displacement with minimal high-quality data. Meanwhile, the proposed architecture adopts focused and progressive model to accurately describe tissues with varied biomechanical properties rather than singular homogeneous model. We made comparison with 4 state-of-the-art (SOTA) methods in simulating multi-component biomechanical deformations of organs with 100 pairs of testing data. Results show that the accuracy of our method is 50% higher than other methods in different validation matrix. And our method can stably simulate the deformations in 0.005 s per frame, which largely improves the computing efficiency.
Read full abstract