Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces the Ultrasound Vision-Language-Action model (US-VLA), which enhances automated ultrasound scanning by integrating clinical semantic goals with real-time feedback to generate precise probe manipulation actions. The approach addresses limitations in existing methods that rely on rigid reward functions and extensive interaction data, thereby improving generalization and stability across various clinical scenarios. Experimental results show that US-VLA performs competitively in ultrasound probe manipulation tasks, suggesting its effectiveness in real-world applications for abdominal examinations.
US-VLA redefines ultrasound scanning by seamlessly combining clinical semantics with real-time feedback, achieving robust performance across diverse scenarios.
Artificial intelligence-assisted ultrasound scanning enhances diagnostic reliability and efficiency by providing real-time guidance for standardized image acquisition and reducing operator dependence. However, existing reinforcement learning and learning-assisted ultrasound scanning methods typically rely on carefully designed reward functions or extensive interaction data, which limits their generalization ability and stability across different devices, patient populations, and complex clinical scenarios. To address these challenges, we propose an ultrasound vision-language-action model (US-VLA) for automated ultrasound scanning that explicitly encodes clinical semantic goals and generates sequential probe manipulation actions under real-time ultrasound feedback. In particular, we first design an ultrasound-aware expert fusion module to jointly integrate ultrasound observations with auxiliary contextual information, enabling semantic ultrasound feedback to effectively guide the scanning process. Then, we construct US-VLA-Data, a real-world dataset covering liver and kidney examinations, which includes five clinically defined standard planes and comprises 320 expert scanning trajectories with approximately 80,000 synchronized timesteps. Extensive experiments demonstrate that US-VLA achieves competitive performance in ultrasound probe manipulation tasks, indicating its effectiveness and promising generalization within the evaluated abdominal ultrasound setting. The source code is available at https://github.com/VMVLab/US-VLA.