New AI framework ASTRA-Net maps hidden peripheral airways for bronchoscopy

Lung cancer is the most diagnosed cancer globally and is one of the leading causes of cancer-related deaths. Early detection of the same can prove to be a boon for survival. However, collecting tissue samples from small tumors located deep within the lungs remains a big challenge. As many of these lesions are found in the peripheral regions of the lungs, physicians must navigate through an intricate network of tiny branching airways to reach the target.

In these cases, clinicians rely on lung navigation systems that use three-dimensional airway maps reconstructed from computed tomography (CT) scans to guide instruments toward suspicious lesions. Eventually, the accuracy of these navigation systems depends heavily on the completeness of the airway maps to guide them. Creating these maps is again difficult, as the smallest peripheral airways are extremely thin and difficult to distinguish from the surrounding tissues. This restriction raises a concern that if artificial intelligence (AI) systems are trained using incomplete information, they may also overlook clinically important airways.

To address this challenge, researchers from Pusan National University in South Korea developed ASTRA-Net (Anatomical Segmentation with Tree-aware Refinement Attention), an AI framework designed to identify previously overlooked airway branches and generate more complete airway maps. The study was led by Dr. MinWoo Kim from the School of Biomedical Convergence Engineering, Pusan National University, and Dr. Hee Yun Seol from the Division of Pulmonary and Critical Care Medicine, Research Institute for Convergence of Biomedical Science and Technology, Pusan National University Yangsan Hospital, in collaboration with other researchers. This paper was made available online on March 9, 2026, and was published in Volume 45, Issue 6 of the journal IEEE Transactions on Medical Imaging on June 01, 2026.

Unlike conventional models that may be limited by incomplete training annotations, ASTRA-Net was designed to identify anatomically plausible airway structures that may have been omitted from the original labels.

"ASTRANET is not simply another airway segmentation model. It was specifically designed to identify peripheral airways that may have been overlooked during manual annotation, thereby helping to create a more complete roadmap for bronchoscopy," explains Dr. Kim.

ASTRA-Net utilizes a multi-stage deep learning architecture like those employed in advanced medical image analysis. A part of this network learns the overall structure of the lungs, whereas another refines regions where airway boundaries are unclear. An additional attention mechanism helps the model focus on regions where small peripheral airways are difficult to distinguish or may have been omitted from the annotations. This allows the reconstruction of airway branches that are often missing from CT annotations. By incorporating anatomical clues from the surrounding lung tissues and blood vessels, which often run alongside the airways, the system can infer the continuity of previously unrecognized airway pathways.

Researchers also evaluated the framework using multiple datasets as well as clinical CT scans obtained from Pusan National University Yangsan Hospital. This model demonstrated strong performance in identifying fine peripheral airways and remained robust even when CT scans varied in image quality and slice thickness. Additionally, expert review of the model predictions showed that some structures initially counted as false positives were genuine airway branches that had been omitted from the original annotations.

"By providing physicians with a more complete airway roadmap, we hope this technology will improve navigation to difficult-to-reach lung lesions and support the development of future AI-assisted and robotic bronchoscopy systems," concludes Dr. Kim.