World Model AI guides robotic ultrasound with 70% success on carotid scans

A research team led by Siqi Fan and Mingcong Chen has introduced a novel action-conditioned world model framework for autonomous robotic ultrasound guidance, specifically targeting neck scans of the carotid artery and thyroid. The core challenge in robotic ultrasound is that collecting high-quality probe-motion trajectories for training is labor-intensive, and building explicit simulators is difficult due to complex interactions like contact, tissue deformation, and view-dependent acoustic artifacts. The team addresses this with a two-stage model-based learning pipeline. First, a latent conditional diffusion world model is trained to predict future ultrasound observations given recent context frames, probe motions, and time offsets. This world model learns the underlying dynamics of ultrasound imaging without needing a hand-crafted simulator. Second, a goal-conditioned temporal transformer is trained to predict ordered probe motions and fine-tuned using rewards derived from the frozen world model, enabling goal-directed navigation.

In experiments using a self-collected dataset and real-world closed-loop tests, the framework achieved success rates of 70.0% for carotid guidance and 65.0% for thyroid guidance. These results were obtained with limited training trajectories, showcasing the world model's ability to preserve action-dependent anatomical structures during target-directed scans. The work highlights the potential of learned ultrasound dynamics for training goal-directed robotic probe navigation, reducing reliance on extensive human-collected demonstrations. This approach could significantly lower the barrier for deploying autonomous ultrasound robots in clinical settings, and the methodology may extend to other medical imaging modalities where simulation is challenging.