AutoPlan: Autonomous Surgical Planning Framework for Neurosurgery Interventions With Steerable Needles

BACKGROUND: Accurate path planning for minimally invasive neurosurgery is necessary to navigate complex anatomy that contains critical structures, such as blood vessels, ventricles and white-matter tracts. Planning with nonholonomic kinematics is difficult, and typical planners fail to generate kinematically feasible paths. METHODS: AutoPlan is a deep reinforcement learning (DRL) framework that generates a needle trajectory on a magnetic resonance imaging (MRI) patient dataset. AutoPlan implements kinematic control strategies, variable curvature with fixed insertion (VCFI) and variable insertion with fixed curvature (VIFC), with random and selective start-point initialisation. Path planning was performed on 55 volumes using PPO, SAC, DDPG and RRT. RESULTS: 20.29 s. CONCLUSIONS: AutoPlan provides consistent planning while avoiding critical structures, resulting in fast, kinematically safe trajectories.