Robust Trachea Segmentation from CT Imaging Using Fully Automated and Prompt-Based Models

Abstract

Accurate trachea segmentation from computed tomography (CT) is a prerequisite for image-guided airway assessment, toward precision tracheostomy planning and safe endotracheal tube placement. Trachea-specific delineation remains challenging due to elongated geometry, small cross-sectional area, partial-volume effects, motion-trigerred artifacts, and heterogeneous surrounding tissues. Here, we study two complementary segmentation paradigms: a fully automated, self-configuring baseline (nnU-Net) and a prompt-conditioned foundation model (Med-SAM) derived from the Segment Anything Model (SAM). We evaluate both approaches on two heterogeneous regimes: (i) a spatially consistent volumetric dataset (AeroPath) and (ii) a large-scale, slice-based dataset without reliable volumetric continuity (OSIC). We further propose a hybrid inference strategy that enables fully automated prompting by deriving bounding-box prompts for MedSAM from coarse nnU-Net predictions. Results show that dataset structure critically influences reliability, interpretability, and deployment feasibility: volumetric continuity benefits fully automatic segmentation, while prompt-conditioned inference improves ROI-constrained interpretability in slice-based settings but introduces prompt sensitivity. We discuss limitations and outline clinically grounded evaluation and EHR integration directions for precision airway management.