Direct Point-wise Non-Rigid Renal Registration via Simulation-Driven Learning of Biomechanical Material Behavior

Abstract

In robot-assisted partial nephrectomy, the real-time alignment of preoperative 3D virtual models with intraoperative anatomical structures is critical for localising hidden tumours and vasculature. However, non-rigid organ deformation and the partial nature of intraoperative surface observations present significant challenges for traditional registration methods. In this study, we propose a novel, two-stage deep-learning framework for non-rigid kidney registration. Our approach operates directly on raw point clouds to preserve fine-grained geometric details. The architecture employs a PointPWC-based pyramid module to estimate coarse-to-fine surface displacements from intraoperative sensor data. Subsequently, a Graph Neural Network utilises these surface flows to infer the internal deformation of the entire organ. To bridge the gap between deep learning and biomechanical reality, the network was trained on a diverse dataset of organ states generated via finite element method simulations. Experimental results demonstrate that the proposed method effectively reconciles geometric alignment with physical plausibility, providing a high-fidelity, real-time registration solution suitable for intraoperative surgical navigation.