Vision-Based Detection and Real-Time Adaptive Control Schemes for Autonomous Ultrasound Scanning Robots

An autonomous ultrasound scanning robot can automatically contact the human body and perform the scanning task at the designated part, and hence, its deployment can alleviate the increasing shortage of professional manpower in recent years. Such a robot commonly employs two types of visual feedback to finalize the task, that is, the RGB vision in the pre-contact phase and the ultrasound imaging in the scan phase. In this paper, the techniques of deep vision detection and real-time adaptive control are introduced in both phases, respectively, to deal with problems due to the uncalibrated and varying relationship between the robot and the human subject. Specifically, the constructed detection network can guide the robot to approach the desired scan position in the presence of unstructured features and moving human limbs, and the proposed control scheme can drive the robot to track the ultrasound image feature subject to unknown limb model parameters. In addition, both techniques function without prior calibration, even when a new human subject is to be scanned. Experimental results are presented to validate the proposed methods in an established robotic ultrasound scanning platform.