TL;DR
This paper introduces a novel anatomy-aware framework for robotic ultrasound (US) systems to automatically guide the US probe to the correct imaging plane for lung ultrasound (LUS). By combining deep learning for anatomical feature detection with image-based visual servoing and a specialized end-effector, the system achieves precise probe positioning, crucial for consistent and reliable diagnostic imaging across different patients.
What Problem Does This Paper Address?
Acquiring diagnostic-quality ultrasound images, particularly for lung ultrasound (LUS), relies heavily on the sonographer's skill to position the probe correctly to achieve a "standardized imaging plane" (SIP). This operator-dependent process leads to inconsistencies in diagnostic outcomes. Robotic ultrasound systems (RUSS) offer a solution by automating probe manipulation, but a significant challenge remains in extracting reliable anatomical features from US images to guide the robot. Existing methods struggle with the inherent noise, artifacts, and inter-patient variability in US images, making it difficult to generalize probe navigation strategies. This paper aims to bridge this gap by developing a robust, anatomy-aware SIP navigation framework for RUSS, specifically for LUS, that can adapt to patient variations and ensure consistent image quality.
Methods
The proposed framework integrates several key components:
- Anatomy-Aware Feature Extraction: A Convolutional Neural Network (CNN) called LUS-Net, based on the U-Net architecture, is used to simultaneously detect and segment anatomical landmarks (pleural line and rib shadows) in real-time LUS images. This network was trained on both phantom and human LUS datasets.
- Anatomical Feature Map (AFM): To overcome the challenges of inter-patient variability and potential inaccuracies in segmentation, an AFM is generated. This map represents a generic location mapping of detected anatomical features from the current US image to a template image. It focuses on the geometric configuration of features rather than absolute pixel locations, making it robust to variations.
- Image-Based Visual Servoing (IBVS): The AFM-derived visual features are used to drive an IBVS controller. This controller calculates the necessary probe motion (in-plane translation and rotation) to align the detected anatomical features with their corresponding locations in the template image, thereby guiding the probe to the SIP.
- Active-Sensing End-Effector (A-SEE): The system incorporates a custom end-effector that uses onboard sensors to continuously adjust the probe's orientation. This ensures the probe remains perpendicular to the patient's body surface, optimizing US signal quality and compensating for deviations caused by the patient's anatomy or movement.
- Contact Force Regulation: A velocity-based PD controller is employed to maintain a constant, safe contact force between the probe and the patient's skin, preventing discomfort and ensuring consistent acoustic coupling.
- Probe Motion Fusion: The movements calculated by the IBVS controller for anatomical alignment and the A-SEE for probe normal positioning are fused to generate the final, comprehensive probe velocity command for the robot.
Datasets
- Phantom Dataset: LUS images were collected from a lung-mimicking phantom (COVID-19 Live Lung Ultrasound Simulator) to train the LUS-Net model.
- Human Subject Dataset: LUS images were collected from five male volunteers for validation. These images were not used for training the LUS-Net model to assess its generalizability. The LUS-Net was trained on a combination of phantom and human data, with a leave-one-out (LOO) approach used for human data to evaluate performance across different subjects.
Key Insights and Ideas
- Anatomy-Awareness is Crucial: Simply relying on image intensity or confidence maps is insufficient for robust US navigation. Explicitly identifying and using anatomical landmarks (like the pleural line and rib shadows in LUS) provides a more reliable basis for guiding the probe.
- Template Matching with Geometric Configuration: The AFM approach is a clever way to handle inter-patient variability. By focusing on the relative positions of anatomical features rather than absolute pixel coordinates, the system can adapt to different patient anatomies without needing patient-specific templates.
- Synergy of Deep Learning and Classical Robotics: The paper effectively combines the power of deep learning for feature detection with the precision of classical robotics control (IBVS) and specialized hardware (A-SEE). This hybrid approach leverages the strengths of each domain.
- "Last Centimeter" Guidance: The framework is designed to work when the probe is already in close proximity to the SIP, focusing on the critical final adjustments needed for optimal image acquisition. This makes it a practical solution for enhancing existing RUSS.
Key Results
- High Navigation Precision: The system achieved positioning errors of under 2 mm in translation and under 2 degrees in rotation, consistently locating the probe at the SIP.
- Improved Image Similarity: During phantom experiments, mean US image similarity improved by up to 92.4% (NSSD metric). For human subjects, the average increase was 90.1% (NSSD metric).
- Robustness to Patient Variability: The framework demonstrated effectiveness on five human subjects with varying body mass indices and chest wall thicknesses, showing adaptability to different anatomies.
- Efficient Navigation: The SIP navigation process was completed within approximately 3-5 seconds for both phantom and human experiments.
- Accurate Contact Force Control: The contact force error was kept low (e.g., 0.31 ± 0.32 N for human subjects), ensuring patient comfort and safety.
Limitations and Future Work
- Limited Human Subject Diversity: The human subject study involved only five male volunteers without lung conditions. Future work should expand the participant pool to include females and individuals with various lung pathologies to further validate the system's generalizability and clinical feasibility.
- Phantom vs. Human Rib Visibility: The rib shadow (RS) feature was less identifiable in the phantom compared to human subjects due to differences in acoustic properties. While the system still performed well, this highlights the importance of realistic phantoms or robust feature detection for diverse anatomical structures.
- Generalizability to Other US Modalities: While the AFM generation and IBVS servoing are formulated generically, the LUS-Net is specific to lung ultrasound. Future work will involve developing feature detection modules for other US examinations (e.g., liver, thyroid) to extend the framework's applicability.
- Pain Sensitivity: While no discomfort was reported, the authors acknowledge that highly pain-sensitive individuals or those with larger body habitus might require excessive probe pressure, which needs further investigation.
Summary updated