Introduction
Robotic ultrasound (RUSS) systems have the potential to standardize medical imaging and improve diagnostic accessibility. However, their application in thoracic procedures is significantly hindered by the rib cage. Ribs reflect ultrasound waves, creating "acoustic shadows" that obscure vital organs like the heart and liver. To obtain clear diagnostic images, a robotic probe must be precisely positioned within the intercostal space—the narrow gap between ribs.
Achieving this precision requires accurate registration between pre-operative CT scans (used for planning) and real-time ultrasound data. Traditional registration methods often rely on the skin surface as a landmark, but skin is highly deformable and does not accurately reflect the underlying skeletal structure. This paper presents a novel non-rigid registration method that utilizes a "skeleton graph" of rib cartilage to bridge the gap between CT and US data.
Methodology
The proposed pipeline shifts the focus from volatile surface geometry to stable anatomical topology. The process involves five key stages:
- Point Cloud Generation: Rib cartilage is manually annotated in both a template CT scan and a 3D ultrasound sweep of the patient to create initial point clouds.
- Self-Organizing Maps (SOM): To manage noise and varying point densities, the system employs SOM—an unsupervised learning technique. This is used first to unify point cloud sizes and subsequently to extract "key points" that represent the essential structure of the ribs.
- Minimum Spanning Tree (MST): These key points are connected using an MST algorithm to form a tree graph, ensuring all structural elements are connected without redundant loops.
- Path Optimization: Recognizing that ribs are continuous anatomical structures, the researchers optimize the graph by identifying the shortest paths between endpoints and filtering out jagged or non-anatomical connections.
- Non-rigid Registration: Finally, the registration is performed by matching these optimized skeleton graphs. Unlike rigid registration, this approach allows the template model to bend and stretch to accommodate the unique anatomical variations of the specific patient.
Experimental Setup
The study utilized seven human chest CT scans from the public RibSeg dataset to serve as templates and test cases. Real-world ultrasound data was acquired from a healthy 26-year-old male volunteer using a KUKA LBR iiwa robot integrated with an ACUSON Juniper ultrasound system. This setup allowed the researchers to test the algorithm's performance against actual robotic scanning conditions.
Key Insights
Cartilage as a Strategic Landmark
While bone is often difficult to penetrate with ultrasound, rib cartilage has low acoustic impedance. This allows the "pleural line" (the lung surface) to be visible directly beneath it, making cartilage an ideal landmark that is consistently identifiable in both CT and US modalities.
Topology over Geometry
By converting raw point clouds into skeleton graphs, the method prioritizes the topology—the connectivity and order of the ribs—over raw spatial coordinates. This makes the system significantly more robust to inter-patient variation and less dependent on precise initial alignment, a common failure point for standard algorithms like Iterative Closest Point (ICP).
Results and Discussion
The skeleton graph approach demonstrated superior performance compared to industry-standard benchmarks:
- Accuracy: The method achieved a mean distance error of 5.0 ± 1.9 mm across the test cases.
- Comparative Performance: This result significantly outperformed both ICP (8.6 ± 6.7 mm) and the Coherent Point Drift (CPD) algorithm.
- Structural Integrity: The method maintained lower Hausdorff distance results, indicating that the anatomical structure of the ribs remained intact during the non-rigid deformation process, whereas standard non-rigid ICP often resulted in anatomical distortions.
From a clinical perspective, a 5mm error margin is highly effective. Given that the average intercostal space is approximately 15mm and the standard ultrasound probe width is 10mm, this level of accuracy provides the robot with the necessary tolerance to navigate the acoustic windows of the thorax successfully.
Limitations and Future Work
While promising, the current study identifies several areas for further development. The current model does not explicitly account for respiratory motion, which causes the ribs to shift during scanning. Additionally, the reliance on manual annotation of rib cartilage in US images remains a bottleneck; the authors plan to automate this process using deep learning-based segmentation. Finally, future iterations will need to account for soft tissue deformations, such as breast tissue in female patients, which can introduce further complexities in ultrasound imaging.
Summary updated