Xuesong Li, Shuai Chen, Feng Li, Zhongliang Jiang, Nassir Navab, Yuan Bi
The paper introduces SonoGraph-WM, a world model that uses scene graphs to represent anatomy and predict how probe movements affect visualizations during robotic ultrasound navigation. It leverages a Transformer to forecast future anatomical states and poses, enabling anticipatory planning through receding-horizon trajectory simulation. The model is trained on CT-derived data with simulated probe trajectories, reducing dependence on annotated ultrasound sequences.
Robotic ultrasound systems often lack the ability to anticipate anatomical changes during probe motion, limiting their autonomy. By modeling anatomy as scene graphs and predicting future views, this approach enables more reliable, goal-directed navigation without requiring real-time image synthesis. This advances the feasibility of autonomous ultrasound in clinical settings where precise, adaptive probe control is critical.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.