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

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Pulmonary Medicine

Background:

• Accurate segmentation of peripheral airway branches is crucial for bronchoscopic navigation.
• Current deep learning (DL) methods face challenges with incomplete annotations and CT acquisition variability.
• Existing segmentation techniques often lack sensitivity in detecting smaller, peripheral airways.

Purpose of the Study:

• To develop a novel DL framework, ASTRA-Net, for robust airway segmentation from CT scans.
• To improve the segmentation of both annotated and previously unlabeled airway branches.
• To enhance the sensitivity and generalizability of airway segmentation for clinical applications.

Main Methods:

• ASTRA-Net integrates auxiliary anatomical inputs (lung and vessel masks) for structural context.
• An encoder-guided attention (EGA) module refines uncertain boundary regions.
• A centerline- and branch-aware loss weighting scheme prioritizes peripheral airways.
• A resolution-robust post-optimization step enhances sensitivity across varying CT resolutions.

Main Results:

• ASTRA-Net achieved superior tree and branch detection rates (TDR and BDR) compared to existing methods.
• The framework demonstrated competitive overlap scores, indicating high segmentation accuracy.
• Experiments across multiple datasets confirmed ASTRA-Net's strong generalizability.

Conclusions:

• ASTRA-Net offers a significant advancement in airway segmentation for CT imaging.
• The proposed framework shows potential for integration into image-guided bronchoscopic procedures.
• Improved airway segmentation accuracy can enhance diagnostic capabilities and therapeutic interventions.