AI uncovers overlooked lung airways to improve bronchoscopy maps

Lung cancer is a prevalent and lethal disease worldwide, with early detection significantly impacting survival rates. However, accessing tissue samples from small tumors deep within the lungs poses a considerable challenge due to the intricate nature of the lung's complex network of tiny branching airways. Physicians often rely on lung navigation systems that utilize three-dimensional airway maps, constructed from computed tomography (CT) scans, to guide instruments toward suspicious lesions.

These maps, while crucial, are often incomplete because the smallest peripheral airways are exceedingly thin and challenging to differentiate from surrounding tissues. This issue has sparked concerns about the potential for artificial intelligence (AI) systems, trained with incomplete data, to overlook clinically significant airways.

Researchers from Pusan National University in South Korea have developed ASTRA-Net (Anatomical Segmentation with Tree-aware Refinement Attention), an AI framework designed to detect previously overlooked airway branches and generate more comprehensive airway maps. Presented in IEEE Transactions on Medical Imaging, the study was spearheaded by Dr. MinWoo Kim from the School of Biomedical Convergence Engineering and Dr. Hee Yun Seol from the Division of Pulmonary and Critical Care Medicine, collaborating with other experts.

Unlike standard models, ASTRA-Net is specifically engineered to uncover peripheral airways that may have been missed during manual annotation, thus enhancing the creation of a more thorough roadmap for bronchoscopy. Utilizing a multistage deep-learning architecture, ASTRA-Net initially learns the general structure of the lungs, followed by refining regions with ambiguous airway boundaries.

An added attention mechanism enables the model to concentrate on regions where minute peripheral airways are challenging to distinguish or have been omitted from annotations. By integrating anatomical cues from adjacent lung tissues and blood vessels, which frequently parallel the airways, the system can deduce the continuity of airway pathways that were previously unnoticed.

The researchers validated ASTRA-Net using various datasets and clinical CT scans from Pusan National University Yangsan Hospital. The model exhibited strong performance in identifying fine peripheral airways and maintained robustness even when CT scans differed in quality and slice thickness. Furthermore, expert evaluation of the model's predictions revealed that certain structures initially classified as false positives were indeed genuine airway branches omitted from the original annotations.

By equipping physicians with a more complete airway roadmap, ASTRA-Net aims to improve navigation to challenging-to-reach lung lesions and facilitate the development of future AI-assisted and robotic bronchoscopy systems.