New AI Framework ASTRA-Net Maps Hidden Peripheral Lung Airways

The challenge of early lung cancer detection lies in the anatomy of the organ itself. Physicians must guide instruments through a complex, microscopic branching system to reach tumors buried in the peripheral regions. Traditional navigation systems rely on CT scans, but thin airways are frequently indistinguishable from surrounding tissue, leading to incomplete digital maps. If an AI model is trained on these incomplete datasets, it inherently repeats the same diagnostic blind spots.

ASTRA-Net, or Anatomical Segmentation with Tree-aware Refinement Attention, moves beyond standard segmentation. Led by Dr. MinWoo Kim and Dr. Hee Yun Seol, the research team published their findings in IEEE Transactions on Medical Imaging. Unlike conventional models, the framework uses a multi-stage architecture that leverages anatomical clues from nearby blood vessels to infer the continuity of unrecognized pathways. By focusing on regions where boundaries are unclear, the system identifies branches that were excluded from the original human-annotated training data.

Clinical evaluations at Pusan National University Yangsan Hospital confirm the system’s robustness across varying scan qualities. During testing, structures initially flagged as false positives were identified by experts as genuine, previously unmapped airway branches. This refinement offers a more precise roadmap for clinicians, with potential applications for future robotic-assisted bronchoscopy and earlier intervention for hard-to-reach lesions.