Technical Summary: Learning-based Hierarchical Tracheal Anatomy Understanding from Sparse Surgical Demonstration Annotations for Ultrasound Robots

Problem Statement

Tracheostomy, a critical procedure for patients requiring mechanical ventilation, relies heavily on the precise localization of the tracheal incision site (typically between the second and third tracheal rings). Conventional manual palpation is subjective, operator-dependent, and unreliable in patients with complex neck anatomy (e.g., obesity or edema). While ultrasound (US) offers a non-invasive alternative, its diagnostic utility remains constrained by operator skill, leading to inter-observer variability and image artifacts.

Robotic Ultrasound (RUS) systems have been developed to automate probe movement, yet current iterations often lack high-level semantic image interpretation. They rely on predefined trajectories or basic force feedback, failing to autonomously adapt to anatomical variations or identify optimal imaging windows. Furthermore, existing deep learning approaches for medical segmentation, such as U-Net, often suffer from anatomical fragmentation and performance degradation when applied to unseen domains or complex anatomical variations. The core challenge addressed is achieving robust, real-time, hierarchical understanding of tracheal anatomy from sparse surgical annotations to enable autonomous robotic assistance.

Methodology

1. Hierarchical Anatomical Reasoning

The authors propose a framework that resolves the ambiguity of tracheal structures (Thyroid Cartilage, Cricoid Cartilage, and Tracheal Rings), which share nearly identical hypoechoic properties. Instead of classifying tissue bodies directly, the system tracks acoustic shadows cast beneath these structures.

• Two-Stage Logic: The system first detects the Thyroid Cartilage (TC) as a global anchor. Based on the TC's proximity to the image frame boundaries, it establishes a directional tracking vector (left-to-right or right-to-left).
• Sequential Identification: Once the TC is localized, the system identifies the Cricoid Cartilage (CC) and sequences the downstream tracheal rings (T1...Tn). If the TC is not detected, the system defaults to a horizontal sort of visible acoustic shadows, mapping them to an ascending sequence of tracheal rings.
• Data Constraints: To ensure topological consistency, data acquisition is restricted to the longitudinal plane with the transducer aligned to the tracheal midline.

2. Hybrid Two-Stage Perception Pipeline

The proposed architecture decouples object detection from pixel-level segmentation:

• Stage 1: Localization (YOLO): A lightweight object detector (specifically YOLOv8n, compared against v10n and v11n) processes the full ultrasound frame to predict bounding boxes for the Thyroid Cartilage and a generalized Cartilage Ring class. This stage filters peripheral background noise and defines the Region of Interest (ROI).
• Stage 2: Segmentation (SAM2): The Segment Anything Model 2 (SAM2) acts as a foundation model decoder. Unlike traditional pipelines that require cropping or shrinking input images, SAM2 accepts the raw bounding box coordinates from YOLO as "prompts."
  • Streaming Memory: SAM2 utilizes a transformer-based streaming memory architecture, treating video sequences as a continuous flux. This preserves temporal consistency during transducer movement, resolving fuzzy tissue walls that frame-by-frame models often misinterpret.
  • Sparse Prompting: By constraining mask decoding to targeted, sparse ROIs defined by YOLO, the system avoids the computational overhead of dense, full-canvas attention calculations.

3. Sparse Annotation and Training Strategy

To address the scarcity of labeled medical data, the authors employed a sample-efficient, sparse annotation protocol:

• Data Curation: Out of 1,205 total video frames, only 90 frames (approx. 7.47%) were manually labeled. Labels were generated via fixed-interval temporal downsampling (e.g., every 4th to 25th frame) based on visual complexity.
• Hybrid Training: The model was trained on a hybrid dataset combining curated in-house laboratory data (controlled environment) and unconstrained web-sourced sequences (generalization environment). This strategy bridges the domain gap, ensuring clinical robustness against varied acoustic settings, probe frequencies, and artifacts.

Key Contributions

1. Integrated Robotic System: Construction of a YOLOv8n-SAM2 integrated system for automatic tracheal anatomy recognition, enabling objective, low-latency clinical localization to replace manual operation.
2. Architectural Optimization: Systematic testing of multiple YOLO versions to confirm YOLOv8n as the optimal backbone for balancing accuracy and speed when paired with SAM2. The paper provides a technical reference by cross-comparing this hybrid approach against U-Net baselines.
3. Robust Hierarchical Detection: A method leveraging clinical constraints (acoustic shadow tracking) to differentiate cartilage structures and achieve robust landmark detection under complex anatomical conditions and domain shifts.

Experimental Results

The framework was evaluated across two domains: a Controlled Domain (curated lab data) and a Generalization Domain (unconstrained web-sourced data).

• Performance Metrics: The YOLOv8n+SAM2 framework trained on the hybrid dataset achieved a consistent Mean Dice Similarity Coefficient (DSC) of 0.777 across both domains.
• Generalization Capability: Models trained exclusively on in-house data suffered catastrophic failure in the generalization domain (DSC ≤ 0.013), often misidentifying background noise as anatomy. The hybrid training strategy was essential for maintaining performance.
• Comparison with Baselines:
  • U-Net Baselines: Standalone U-Net models achieved lower generalization DSC (≤ 0.494) and frequently merged adjacent tracheal rings or produced fragmented masks.
  • YOLO+U-Net: While showing some adaptability, this configuration suffered from computational bottlenecks and sparse, fragmented outputs (Mean DSC 0.339–0.601).
• Efficiency: The proposed framework achieved a throughput of 6.92 FPS in the generalization domain. This is vital for closed-loop robotic teleoperation, significantly outperforming the latency of the YOLO+U-Net pipeline (approx. 2.14 FPS).

Significance and Claims

The paper claims that this study establishes a scalable foundation for standardized, autonomous surgical assistance. By coupling lightweight localization with foundation-scale visual models, the framework successfully navigates the variability of real-world ultrasound data.

Key claims include:

• Robustness: The decoupled architecture effectively balances generalization, precision, and efficiency, overcoming the trade-offs inherent in medical mechatronics.
• Safety and Precision: The system provides a reliable technical basis for more accurate robotic-assisted tracheostomy, reducing reliance on subjective manual palpation.
• Feasibility of Sparse Learning: The study confirms that a robust hierarchical understanding of tracheal anatomy can be derived from sparse datasets (approx. 7.5% labeling) when using a hybrid training strategy.

Limitations Acknowledged by the Authors:
The authors modestly note that while the current processing speed (~6.9 FPS) is sufficient for visual guidance, it operates below the 500 Hz–1 kHz benchmark required for mechanical stability and haptic feedback in closed-loop control systems. Additionally, the framework's robustness is currently tied to the specific dataset used, and broader generalizability across diverse clinical populations requires further data diversification. Future work aims to decouple the visual perception module from high-frequency haptic loops.