Reproduce “Goal-conditioned reinforcement learning for ultrasound navigation guidance” on OpenResearch

About this paper

Abdoul Aziz Amadou, Vivek Singh, Florin C. Ghesu, Young-Ho Kim, Laura Stanciulescu, Harshitha P. Sai, Puneet Sharma, Alistair Young and 2 more

Researchers from King's College London and Siemens Healthineers developed a goal-conditioned reinforcement learning framework for Transesophageal Echocardiography (TEE) that guides probe navigation to arbitrary views. The system, leveraging novel contrastive patient batching and data augmentation, achieves an average angular error of 9.36 degrees and positional error of 6.56 mm for standard views, demonstrating strong generalization across patient anatomies and to unseen interventional views.

Key insights

• A single, generalist GCRL model can effectively learn to navigate an ultrasound transducer to arbitrary, visually-defined views, overcoming the 'one model per view' limitation of prior approaches.
• The proposed Contrastive Patient Batching (CPB) and data-augmented contrastive loss are critical for improving model generalization across diverse patient anatomies and unseen target views.
• High-fidelity, physics-based simulation environments are a viable and effective means for generating the extensive data required to train complex deep reinforcement learning agents for medical imaging tasks.

Summary by alphaXiv.