Bronchoscope Tip Localization With Learned Airway Constraints and an Open Source Simulator

Abstract

Accurate bronchoscope localization is challenging because airway images contain limited texture and similar views across branches. We present BronchoLoc, a learning‐based framework that combines monocular video with a preoperative CT centerline graph to estimate the 3D bronchoscope tip position. Local visual tracking and global sequence context constrain predictions to the airway anatomy. The accompanying open‐source BronchoSim simulator generates synchronized images, depth, and trajectories (Traj) for training and evaluation. Averaged across Traj, localization RMSE was 5.65 mm in simulation with 82.8% branch selection accuracy and 8.72 mm in the physical phantom. Together, BronchoLoc and BronchoSim provide a framework for developing and evaluating image‐based airway localization.