An imitation learning framework anchored on anatomy obtains standard kidney, vesicle and liver cuts from 16 participants never seen in training.
Ultrasound is highly dependent on the operator, hence the interest in robots capable of reproducing the gestures of an expert on a constant basis. The researchers propose to start by identifying, from external images, an anatomical starting point specific to the examination; the initial placement of the probe then serves as a reference point for the entire scan. A flow-based policy then generates the movements by combining the position of the probe relative to this marker, the ultrasound images, a camera attached to the probe and the contact forces. The demonstrations were collected without a robot, using a portable platform used during freehand examinations.
Out of 16 novel participants, the system achieved the correct cut 96% of the time for the kidney, 94% for the gallbladder and 88% for the liver, and succeeded in 94% of radial artery scans. Compared to a refined π0 model, gains reached 10 to 42 points on abdominal examinations, with less time to reach the target and lower contact forces.
With automatic placement, end-to-end trials succeed 90-100% of the time, and policies work without retraining on a second robot, with 83-100% success.
Read the full article on the original source
Go to the source →