Autonomous Robotic Liver Ultrasound via Vision-Guided Initialization and Imitation Learning

Abstract

Liver disease is a leading cause of mortality worldwide. Ultrasound (US) examination is a primary modality for monitoring and assessing the risk of liver disease. Robotic ultrasound system (RUSS) has emerged as a promising approach to reduce dependence on sonographers and to improve the standardization and reproducibility of US examination. To advance liver US examination, this study proposes a fully autonomous two-stage methodology for robotic liver US examination, integrating vision-guided initial positioning and imitation learning (IL). In the initial positioning stage, the probe autonomously establishes initial contact with patients using RGB-D images and determines the transition timing to the IL stage based on real-time US images. In the IL stage, scanning skills–represented by robot trajectories in Cartesian space and corresponding US image state trajectories–are learned and reproduced using an enhanced IL method, Adaptive Ensemble Transfer Bayesian Interaction Primitive (AET-BIP). Compared with its original formulation, AET-BIP reduces fluctuations in scanning phase velocity, thereby improving stability. To evaluate repeatability, the robot’s end-of-scan poses were measured under varying initial conditions in the IL stage, yielding position and orientation repeatability errors generally below 2.50 mm and 0.020 rad, respectively. Comprehensive experiments incorporating both stages validate the feasibility and adaptability of the proposed methodology, highlighting its potential for standardized and reliable clinical liver US examination.