Autonomous path planning for intercostal robotic ultrasound imaging using reinforcement learning.

Area of Science:

• Robotics and Artificial Intelligence in Medical Imaging

Background:

• Ultrasound (US) imaging is crucial for clinical diagnosis but suffers from operator variability.
• Robotic systems offer enhanced precision and repeatability for US examinations.
• Current autonomous US navigation often lacks 3D anatomical context, hindering path planning in complex regions.

Purpose of the Study:

• To develop a novel reinforcement learning (RL) approach for autonomous intercostal US scanning path planning.
• To integrate 3D anatomical knowledge from computed tomography (CT) templates into the RL decision-making process.
• To enable efficient and non-shadowed US trajectory planning in anatomically challenging areas.

Main Methods:

• A virtual environment was created using CT templates with simulated tumors for training.
• A reinforcement learning framework was employed, utilizing 3D state representations.
• Task-specific state representations and reward functions were designed to optimize scanning and mitigate artifacts like acoustic attenuation and shadows.

Main Results:

• The proposed RL framework successfully planned non-shadowed US scanning trajectories.
• The approach demonstrated efficiency in planning paths for single or multiple scanning targets on unseen patient models.
• The method effectively navigated areas with limited acoustic access.

Conclusions:

• The study presents an effective RL-based solution for autonomous US path planning, particularly in complex anatomical regions.
• Integrating 3D anatomical information significantly improves US scanning path planning accuracy and reliability.
• This work lays the foundation for future development in robotic ultrasound execution and control.