Robot Learning Ultrasound Scanning using Neural Dynamic Movement Primitives

In the post-COVID era, robot manipulators have been utilized in a variety of medical procedures, including diagnosis, surgery, rehabilitation, etc. Among these procedures, Robotic Ultrasound Scanning (RUS) has gained a lot of attention as it is the most commonly used imaging modality and manual procedure requires sonographers to come in direct physical contact with patients. However, robotic automation of ultrasound is quite challenging, as the human body exhibits considerable variability in physiological and anatomical conditions. Sonographers use their diagnostic skill from prior medical education and training to conduct inter-patient procedures. Thus, transferring this skill from expert sonographer to robot is essential for realizing the RUS as per clinical protocols.

In this work, we present Neural Dynamic Movement Primitives (N-DMP) for adapting the DMPs to RUS skill learning. DMPs are a commonly used skill representation model. However, they have not been explored beyond skill cloning. Moreover, they have been modeled using single sensory information. The proposed N-DMP allows learning from multi-modal, high-dimensional, and multiple demonstrations of sensory information, including kinematics (probe position and orientation), dynamics (probe forces), and visual (ultrasound images) data. This framework paves the way for encoding the domain expertise of expert sonographers in the form of probing maneuvers and ultrasound image analysis for inter-patient RUS.