Segmented Trajectory Optimization of Flexible Needles Based on Deep Reinforcement Learning

Abstract

Minimally invasive liver tumor ablation relies heavily on safe and accurate flexible needle trajectory planning. Traditional methods optimize the entire path uniformly and ignore distinct clinical requirements of extrahepatic and intrahepatic tissues. They also suffer from premature convergence in complex high-dimensional spaces. To address these issues, this paper proposes a segmented trajectory optimization framework based on deep reinforcement learning enhanced particle swarm optimization (PSO). We introduce a deep deterministic policy gradient (DDPG) agent to dynamically adjust PSO parameters, which improves global search ability and convergence performance. Moreover, a tissue-layer-aware segmented strategy is designed. The puncture path is divided into extrahepatic and intrahepatic segments with differentiated cost functions. The framework minimizes non-target tissue damage and drives intrahepatic length to a clinically optimal range. Simulation experiments on 3D patient-specific anatomical models show that the proposed method outperforms GA-PSO and pure DDPG in convergence, path safety and clinical compliance. It effectively balances needle fixation stability and surgical safety. This work provides a clinically adaptive computational tool for patient-specific preoperative planning. Limitations include lack of physical and clinical validation and simplified dynamic environment modeling. Future work will focus on real-world verification and complex dynamic tissue interaction.