Class-Aware Cartilage Segmentation for Autonomous US-CT Registration in Robotic Intercostal Ultrasound Imaging

TL;DR

The paper tackles autonomous thoracic ultrasound planning under patient variability by introducing CUS-Net, a class-aware cartilage segmentation framework, coupled with geometry-constrained VAE post-processing. It pairs coarse segmentation, bone classification, and CAM-guided fine segmentation to produce precise cartilage bone masks, which feed a dense skeleton graph-based non-rigid registration to transfer a CT-planned intercostal path to individual patients. Across a multi-step pipeline, it demonstrates Dice improvements to cartilage segmentation (up to $0.88\pm0.09$) and achieves precise path transfer with Euclidean errors around $2.2\pm1.1$ mm, outperforming classic registration baselines. This approach enables robust, automated intercostal US for liver and thoracic interventions and offers a foundation for template-based autonomous robotic ultrasound with potential extensions to other thoracic applications and continuous path planning.

Abstract

Ultrasound imaging has been widely used in clinical examinations owing to the advantages of being portable, real-time, and radiation-free. Considering the potential of extensive deployment of autonomous examination systems in hospitals, robotic US imaging has attracted increased attention. However, due to the inter-patient variations, it is still challenging to have an optimal path for each patient, particularly for thoracic applications with limited acoustic windows, e.g., intercostal liver imaging. To address this problem, a class-aware cartilage bone segmentation network with geometry-constraint post-processing is presented to capture patient-specific rib skeletons. Then, a dense skeleton graph-based non-rigid registration is presented to map the intercostal scanning path from a generic template to individual patients. By explicitly considering the high-acoustic impedance bone structures, the transferred scanning path can be precisely located in the intercostal space, enhancing the visibility of internal organs by reducing the acoustic shadow. To evaluate the proposed approach, the final path mapping performance is validated on five distinct CTs and two volunteer US data, resulting in ten pairs of CT-US combinations. Results demonstrate that the proposed graph-based registration method can robustly and precisely map the path from CT template to individual patients (Euclidean error: $2.21\pm1.11~mm$).

Figures (6)

• Figure 1: (a) Illustration of US liver scan from intercostal space and three types of thorax bones: sternum, rib and costal cartilage. (b), (c) and (d) are the representative US images acquired on the sternum, rib and cartilage, respectively. They have distinct anatomical features on US images.
• Figure 2: The proposed class-aware cartilage bone segmentation architecture. The CUS-Net consists of four distinct modules: coarse segmentation, classification, fine segmentation, and boundary-constraint VAE-based post-processing. First, a coarse segmentation network is employed to generate region proposals for the target anatomy. Subsequently, a classification module is utilized to automatically differentiate between cartilage, rib, and sternum regions. Leveraging the Class Activation Maps (CAM) generated by the classification module, a fine segmentation process is conducted to improve segmentation accuracy. Finally, a boundary-constrained VAE-based post-processing module is applied to refine the shape accuracy of the cartilage bone, ensuring robust inputs for registration.
• Figure 3: Illustration of coarse alignment between CT and US point clouds. The US point cloud is generated based on the autonomous segmented cartilage US images from volunteers and the paired robotic tracking information. The CT point cloud was generated through manual annotation of the CT chest volume. By precisely segmenting the sternum and individual cartilage branches in both the CT template and patient-specific US point clouds, the two point sets can be coarsely aligned by matching the sternum.
• Figure 4: The illustration of the fine alignment of CT and US skeleton point clouds using the SOM algorithm based on the geodesic distance. The graph node correspondences can be obtained based on the optimized $\textbf{G}_{ct}$ and $\textbf{G}_{us}$.
• Figure 5: The illustration of the VAE-based boundary-constraint postprocessing results in various cases.