Discovery of Peripheral Airway Beyond Incomplete CT Annotations for Navigational Bronchoscopy.

Abstract

Sensitive reconstruction of peripheral airway branches is critical for bronchoscopic navigation, yet existing deep learning (DL) methods often struggle with incomplete annotations and variability in CT acquisitions. We propose ASTRA-Net, a DL framework for segmenting both annotated and previously unlabeled airway branches from CT scans. ASTRA-Net integrates four key components: (1) auxiliary anatomical inputs (lung and vessel masks) to provide structural context, (2) an encoder-guided attention (EGA) module to refine boundary regions with high structural uncertainty, (3) a centerline- and branch-aware loss weighting scheme to emphasize clinically important peripheral branches, and (4) a resolution-robust post-optimization step to improve sensitivity under varying slice thickness. Experiments on multiple public and inhouse datasets show that ASTRA-Net achieves the highest tree and branch detection rates (TDR and BDR) while maintaining competitive overlap scores, demonstrating strong generalizability across domains. These results indicate ASTRA-Net's potential as a step toward integrating robust airway segmentation into image-guided bronchoscopic procedures. The source code is available at: https://github.com/pnu-amilab/Airway.