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

Zhongliang Jiang, Yunfeng Kang, Yuan Bi, Xuesong Li, Chenyang Li, Nassir Navab

Introduction

Medical ultrasound (US) is a cornerstone of diagnostic imaging, yet its efficacy is often limited by the presence of bone. Ribs, characterized by high acoustic impedance, reflect the majority of ultrasound waves, creating acoustic shadows that obscure vital organs such as the liver and heart. For autonomous robotic ultrasound systems, navigating these "intercostal spaces" is a significant challenge due to the high variability in patient anatomy. Traditional registration methods, such as those relying on external skin markers, often fail to accurately account for internal skeletal geometry. This work introduces a novel framework for real-time identification and registration of subcutaneous thoracic bone structures to facilitate autonomous intercostal scanning.

Anatomical context and ultrasound signatures of thoracic structures: (a) rib cage model with scanning trajectory, (b) sternum, (c) costal cartilage with pleural line, and (d) rib.

Methodology

The proposed system employs a multi-stage pipeline designed to map a pre-operative CT template onto a patient's live ultrasound scan.

Class-Aware Segmentation (CUS-Net)

The core of the system is CUS-Net, a deep learning architecture that classifies thoracic structures into five categories: cartilage, rib, sternum, transition parts, and background. Unlike generic bone detectors, CUS-Net utilizes a coarse-to-fine strategy. A classification module generates Class Activation Maps (CAMs), which guide a subsequent segmentation module to focus on relevant anatomical regions.

The CUS-Net architecture, featuring coarse segmentation, class-aware classification, fine segmentation with CBAM, and VAE-based post-processing.

VAE-Based Post-Processing

To ensure anatomical plausibility, the system incorporates a Variational Autoencoder (VAE) trained on valid bone morphologies. If the initial segmentation produces fragmented or irregular shapes, the VAE projects the mask into a learned manifold space to retrieve the nearest anatomically correct representation.

Dense Skeleton Graph Registration

Following segmentation, the system constructs a graph of the identified skeletal structures. A modified Self-Organizing Map (SOM) algorithm performs non-rigid registration between the CT-derived skeleton and the live US data. This allows the pre-planned scanning path to be dynamically adjusted to the patient's specific rib cage geometry.

Results

The framework was evaluated using a dataset of 8,721 thoracic bone images and five CT volumes.

  • Segmentation Accuracy: CUS-Net achieved a Dice score of 0.88 and an IoU of 0.79 for cartilage segmentation, significantly outperforming standard baseline models (Dice ~0.69).
  • Registration Precision: The system mapped intercostal paths with a mean Euclidean error of 2.21 ± 1.11 mm. This represents a substantial improvement over traditional Iterative Closest Point (ICP) methods, which exhibited errors exceeding 13 mm.
  • Real-Time Performance: The inference pipeline operates in approximately 149 ms, supporting near real-time robotic adjustments during scanning procedures.

Discussion and Conclusion

A key finding of this research is the utility of costal cartilage as a landmark. Unlike ribs, cartilage allows for the visualization of the underlying pleural line, providing a more reliable feature for registration. The integration of class-awareness and VAE-based shape constraints ensures that the robotic system maintains a high degree of accuracy even in the presence of noisy ultrasound data.

Future work will focus on expanding the dataset to include a more diverse patient population and integrating real-time motion compensation to account for respiratory cycles and probe-induced tissue deformation. This approach holds promise for enhancing the autonomy and precision of robotic ultrasound in thoracic applications.

Summary updated

BibTeX
@article{jiang2025class,
  author = {Zhongliang Jiang and Yunfeng Kang and Yuan Bi and Xuesong Li and Chenyang Li and Nassir Navab},
  title = {{Class-Aware Cartilage Segmentation for Autonomous US-CT Registration in Robotic Intercostal Ultrasound Imaging}},
  journal = {IEEE Trans Autom. Sci. Eng.},
  year = {2025},
  doi = {10.1109/tase.2024.3411784},
}
Figure