BCNet: Bronchus Classification via Structure Guided Representation Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Accurate analysis of the bronchial tree using CT scans is crucial for diagnosing lung and airway diseases.
• Individual variations in bronchial tree topology pose a significant challenge for automated classification systems.

Purpose of the Study:

• To introduce the Bronchus Classification Network (BCNet), a novel framework designed to improve automated bronchus classification.
• To leverage segment-level topological information from point clouds for enhanced voxel-level feature learning.

Main Methods:

• Developed BCNet, a structure-guided framework with two branches: a Point-Voxel Graph Neural Network (PV-GNN) for segment classification and a Convolutional Neural Network (CNN) for voxel labeling.
• Simultaneously trained both branches to learn topology-aware features for a shared backbone.
• Ensured efficient inference by enabling the use of only the CNN branch.

Main Results:

• BCNet significantly outperformed state-of-the-art methods, achieving an over 8.0% improvement in F1-score for bronchus classification.
• The proposed framework demonstrated comparable inference efficiency to baseline CNN models.
• Introduced BronAtlas, an open-access benchmark dataset for bronchus imaging analysis with detailed annotations.

Conclusions:

• BCNet offers a robust and efficient solution for topology-aware bronchus classification in medical imaging.
• The BronAtlas benchmark provides valuable resources for advancing research in lung and airway disease diagnosis.
• The developed framework holds promise for improving the accuracy and efficiency of clinical diagnostic tools.