Reproduce “Autonomous Navigation of an Ultrasound Probe Towards Standard Scan Planes with Deep Reinforcement Learning” on OpenResearch

About this paper

Keyu Li, Jian Wang, Yangxin Xu, Hao Qin, Dongsheng Liu, Li Liu, Max Q.-H. Meng

Researchers at CUHK and SUSTech developed SonoRL, a deep reinforcement learning framework for autonomous 6-D ultrasound probe navigation towards standard scan planes. The system achieved a 92% success rate with 4.91mm position and 4.65° orientation errors in intra-patient settings, comparable to human intra-observer variability, and significantly enhanced generalization to unseen anatomies by integrating image quality optimization.

Key insights

• Navigation performance is positively correlated with improvements in image quality, suggesting that clear image feedback is essential for successful probe guidance.
• Explicitly rewarding image quality (via confidence maps) significantly improves the DRL agent's generalization ability, especially when adapting to new, unseen patient anatomies.
• A DRL agent can learn to control the full 6-D pose of an ultrasound probe and adhere to physical constraints like patient contact and tilt limits within a realistic simulation.

Summary by alphaXiv.