Pusan National University Launches ASTRA-Net for Enhanced Mapping of Lung Airways

Lung cancer remains the most frequently diagnosed cancer worldwide and is a top contributor to cancer-related fatalities. The key to enhancing survival rates lies in early detection, yet tackling this challenge is complicated when tumors lurk deep within the lungs. A significant number of these lesions are located in the peripheral regions, necessitating specialized navigation through a dense maze of tiny, branching airways to obtain biopsies and other critical interventions.

To address this urgent need, researchers at Pusan National University in South Korea have unveiled ASTRA-Net, an advanced AI framework tailored specifically for improved lung airway identification and mapping. This innovative technology promises to revolutionize the diagnostic process for lung cancer through its ability to reconstruct often-neglected peripheral airways that traditional systems may miss.

The motivation behind creating ASTRA-Net stemmed from the limitations of lung navigation systems that conventionally rely on computer tomography (CT) scans to provide three-dimensional airway maps. The efficacy of these systems is predicated on the completeness and accuracy of the airway maps they employ, which can be challenging to generate due to the exceptionally thin and delicate nature of the peripheral airways. When AI systems are trained on incomplete airway data, there exists a significant risk of overlooking potentially critical anatomical structures.

Led by Dr. MinWoo Kim from the School of Biomedical Convergence Engineering and Dr. Hee Yun Seol, of the Division of Pulmonary and Critical Care Medicine at the Research Institute for Convergence of Biomedical Science and Technology at Pusan National University Yangsan Hospital, this study emphasizes the importance of complete anatomical mapping in medical imaging. Their findings were published in the prestigious IEEE Transactions on Medical Imaging in June 2026.

Unlike conventional models constricted by inadequate training data, ASTRA-Net is meticulously designed to recognize anatomically plausible structures that may have been dismissed in initial CT scan labels. Dr. Kim details the technology, stating, “ASTRA-Net is not merely another airway segmentation model. It was engineered to prioritize the identification of peripheral airways that are frequently overlooked and, thus, enhance the roadmap available for bronchoscopic procedures.”

ASTRA-Net employs a sophisticated multi-stage deep learning architecture typically utilized in advanced medical image analysis. Its framework splits tasks into various segments: one segment captures the broader anatomical structure of the lungs, while others fine-tune areas where airway boundaries are indistinct. Additionally, an innovative attention mechanism enables the model to zero in on regions with unclear or ambiguously labeled small airways. This comprehensive approach facilitates the reconstruction of those airway branches often missing from standard CT annotations.

In an effort to validate the efficacy of ASTRA-Net, researchers conducted extensive evaluations using various datasets and actual clinical CT scans sourced from Pusan National University Yangsan Hospital. The results showcased the model's impressive performance in accurately locating fine peripheral airways, demonstrating reliability across varying image qualities and slice thicknesses. Notably, through expert analysis of the model's predictions, some features originally considered false positives were confirmed as legitimate airway branches, further emphasizing the model's capability to identify previously unmapped structures.

The overarching goal of this innovative technology is to enhance physician capabilities in navigating to challenging-to-reach lung lesions, ultimately leading to improved patient outcomes in lung cancer procedures. Dr. Kim concludes,