Dual-Agent-Based Robotic Ultrasound Path Planning and Interaction Control in External-Vision-Independent Environments

Ultrasound imaging has emerged as a crucial tool for the diagnosis and navigation of spinal diseases. However, high-quality image acquisition heavily relies on experienced sonographers, which restricts its further popularization. In this paper, a robotic system designed for automated spinal ultrasound scanning is proposed. Drawing inspiration from the spinal anatomy and the actions of seasoned sonographers, the system integrates both a deep learning agent and a reinforcement learning agent to collaboratively guide the adjustment of the ultrasound probe in external-vision-independent environments, relying on real-time ultrasound images and contact force. Then, a hybrid force-to-velocity control framework is proposed to ensure proper ultrasound coupling during the scanning process. Experimental results on a phantom and human participants demonstrated that this system can accurately track spinal features (mean error: less than 1 mm) and maintain normal probe orientation (out-of-plane angular error: 1.61 ± 1.1°, in-plane angular error: 1.27 ± 0.9°), resulting in high-quality and reproducible ultrasound images. Overall, our system shows great potential for clinical applications.