Thoracic Cartilage Ultrasound-CT Registration Using Dense Skeleton Graph

Zhongliang Jiang, Chenyang Li, Xuesong Li, Nassir Navab

Introduction

Autonomous robotic ultrasound systems (RUSS) represent a significant advancement in medical imaging, aiming to reduce the operator-dependency and physical strain associated with manual sonography. A critical challenge in deploying these systems for thoracic imaging is the precise translation of scanning paths planned on a pre-operative 3D CT atlas to the patient’s actual anatomy during the procedure. This paper presents a novel registration method that utilizes the thoracic cartilage and sternum as stable skeletal anchors to ensure the ultrasound beam accurately navigates the narrow acoustic windows between ribs.

The Challenge: Navigating the Intercostal Space

The thoracic region presents a unique difficulty for ultrasound imaging due to the high acoustic impedance of the rib cage. Ultrasound waves cannot penetrate bone or cartilage, creating "shadows" that obscure underlying organs like the heart and lungs. Consequently, the ultrasound probe must be positioned and angled precisely through the intercostal spaces—the narrow gaps between the ribs.

Traditional registration methods often rely on the patient's skin surface. However, skin is highly deformable and its position does not strictly correlate with the underlying skeletal structure that defines the acoustic windows. To achieve reliable autonomous scanning, a registration method must explicitly account for the subcutaneous skeleton to ensure the ultrasound beam reaches its intended target without being blocked by the ribs.

Methodology

The researchers proposed a Dense Skeleton Graph approach for non-rigid registration, moving beyond simple point-cloud matching to a topology-aware deformation model. The workflow involves five key stages:

  1. Point Cloud Generation: 3D point clouds of the sternum and costal cartilage are extracted from both the template CT scan and a live ultrasound sweep of the patient.
  2. Coarse Alignment: To initialize the registration, a 2D template matching approach identifies the sternum. By treating the sternum as a rectangular anatomical biomarker, the system achieves a rough alignment between the CT and US coordinate spaces.
  3. Directed Skeleton Graph: The CT template is represented as a directed graph. Unlike a simple collection of points, this graph structure preserves the anatomical connectivity (topology) of the rib cage.
  4. Modified Self-Organizing Map (SOM): This core innovation involves an SOM algorithm that iteratively deforms the CT graph to match the US point cloud. Crucially, the algorithm utilizes geographical distance (distance measured along the graph edges) rather than standard Euclidean distance. This constraint prevents the algorithm from incorrectly "snapping" a node from one rib onto a neighboring rib simply because they are physically close in 3D space.
  5. Weighted Transformation: A weighted formula maps individual points from the CT space to the US space, ensuring local accuracy while maintaining the global continuity of the thoracic anatomy.

Experimental Setup and Results

The study utilized five thoracic CT scans from the publicly available RibSeg dataset to serve as atlases for path planning. Real-world ultrasound data was captured using a KUKA robotic arm equipped with an ACUSON Juniper US system, scanning a healthy volunteer (male, age 26, BMI 19.8).

The proposed graph-based method was benchmarked against three standard techniques: Iterative Closest Point (ICP), non-rigid ICP, and Coherent Point Drift (CPD).

  • Registration Accuracy: The proposed method achieved a Hausdorff distance of 9.48 ± 0.27 mm, significantly outperforming the baseline methods, which produced errors ranging from 14.58 mm to 17.35 mm.
  • Path Transfer Precision: When transferring 20 pre-planned waypoints from the CT atlas to the US space, the proposed method maintained an average Euclidean error of 2.21 ± 1.11 mm. In contrast, the next best method (CPD) resulted in an error of 3.95 mm, while standard ICP reached 4.78 mm.
  • Robustness: The method demonstrated lower standard deviation across different datasets, suggesting it is more resilient to inter-patient anatomical variations than traditional point-cloud approaches.

Discussion and Future Work

By prioritizing the "geographical" relationship between skeletal structures, the Dense Skeleton Graph method ensures that the 4th rib in a CT scan correctly maps to the 4th rib in the ultrasound data, even under deformation. This topology-preserving approach is essential for maintaining the integrity of the intercostal "windows."

Despite these gains, limitations remain. The current model assumes a static patient; future iterations must incorporate compensation for respiratory motion. Additionally, because costal cartilage is relatively flat, the 3D point clouds lack significant depth information in the direction normal to the chest, which can affect orientation mapping. Future research may integrate internal organ biomarkers to provide a more comprehensive three-dimensional registration.

Conclusion

The transition from skin-based to skeleton-based registration marks a vital step forward for autonomous thoracic ultrasound. By leveraging the stability of the rib cage and the topological constraints of a dense skeleton graph, this method provides the precision necessary for robotic systems to "see" through the complex geometry of the human chest.

Summary updated

BibTeX
@inproceedings{jiang2023thoracic,
  author = {Zhongliang Jiang and Chenyang Li and Xuesong Li and Nassir Navab},
  title = {{Thoracic Cartilage Ultrasound-CT Registration Using Dense Skeleton Graph}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341575},
}
Figure