Abstract
Lung ultrasound scanning is essential for diagnosing lung diseases. The scan effectiveness critically depends on both longitudinal and transverse scans through intercostal spaces to reduce rib shadowing interference, as well as maintaining the probe perpendicular to pleura for pathological artifact generation. Achieving this level of scan quality often depends heavily on the experience of doctors. Robotic ultrasound scanning shows promise, but currently lacks a direct path planning method for intercostal scanning, and probe orientation does not consider imaging differences between lungs and solid organs. In this paper, we aim to fully automate two fundamental operations in lung ultrasound scanning: longitudinal and transverse scans. We propose pioneering path planning methods along intercostal spaces and innovative solutions for adaptive probe posture adjustment using real-time pleural line feedback, specifically addressing the unique characteristics of lung ultrasound scanning. This ensures the acquisition of high-quality, diagnostically meaningful ultrasound images. In addition, we develop a robotic lung ultrasound system to validate the proposed methods. Extensive experimental results on two volunteers and a chest phantom confirm the efficacy of our methods, and demonstrate the system’s feasibility in automated lung ultrasound examinations. Our work lays a solid foundation for automated robotic complete lung scanning.
The Challenge of Lung Ultrasound
Lung ultrasound (LUS) is a powerful, radiation-free tool for diagnosing conditions like pneumonia and COVID-19. However, its effectiveness is hindered by the "rib problem." Because ribs block ultrasound waves, clinicians must expertly navigate the probe through narrow intercostal spaces (ICS) to obtain a clear view of the lung surface, known as the pleura.
Furthermore, LUS diagnosis relies on specific artifacts—reverberations called A-lines and B-lines—that only appear clearly when the probe is held perfectly perpendicular to the pleura. Existing robotic ultrasound systems typically follow the "skin normal" (the direction the skin is facing), but the internal pleura is often tilted differently. This research addresses the challenge of automating this highly skilled manual process, moving beyond simple surface-following to intelligent, feature-aware scanning.
Autonomous Robotic Framework
The researchers developed a multi-stage framework to achieve full autonomy using a 6-DOF robotic arm equipped with an ultrasound probe and an RGB-D camera. The process begins with Patient-Specific Path Planning. By identifying landmarks on the patient (the navel and nipples) using the camera, the system maps "ideal" intercostal centerlines from a standard 3D human thorax model onto the specific patient’s body.
To ensure the probe remains correctly positioned during the scan, the system utilizes Real-Time Pleural Line Segmentation. Using nnU-Net, a state-of-the-art deep learning framework, the system identifies the pleural line in the ultrasound feed at 10Hz. This allows the robot to "see" the lung surface in real-time and adjust its posture accordingly.
Adaptive Control and Feedback
The core innovation of this system is the Adaptive Probe Posture Adjustment loop. Instead of relying solely on external sensors, the system uses the ultrasound image itself as a feedback mechanism:
- Translation: If the pleural line drifts toward the edge of the image, the robot shifts along its long axis to re-center it.
- Rotation: If the pleural line appears tilted, the robot rotates to ensure the ultrasound beams hit the pleura at a 90-degree angle.
This is implemented via a two-step workflow. The robot first performs a longitudinal scan (crossing multiple ribs) to "scout" the exact location of the intercostal spaces. It then uses that data to reconstruct a more accurate path for a transverse scan (sliding between the ribs), ensuring a view entirely unobstructed by bone. Throughout this process, an admittance control scheme ensures the robot maintains a steady, safe contact force of approximately 8 Newtons.
Experimental Results
The nnU-Net segmentation model was trained on a custom dataset of 1,168 ultrasound images manually annotated by clinicians. It achieved high accuracy, with a Dice score of 86.17% for longitudinal scans and 85.08% for transverse scans.
Key performance metrics from the robotic trials include:
- Posture Correction: Without adjustment, the probe's angle relative to the pleura deviated by as much as 22°. With online adjustment, the system reduced this deviation to below 5°, significantly improving the visibility of diagnostic A-lines.
- Centering: Translation adjustments kept the target pleural line within a 5mm deviation from the center of the image.
- Safety: The compliant control system successfully limited contact force spikes to below 13N during initial contact, quickly stabilizing to the target 8N.
Conclusion and Future Directions
This work demonstrates the feasibility of fully autonomous lung ultrasound scanning. By treating the ultrasound image as a sensor, the system can overcome the limitations of surface-based path planning. While currently focused on the anterior chest, future work will expand to the lateral and posterior regions and incorporate methods to account for respiratory motion. This system lays a solid foundation for standardized, expert-level lung diagnostics accessible through robotic automation.
Summary updated