BCNet's structure-guided learning lifts bronchus classification F1 by 8%

Diagnosing lung and airway diseases often requires analyzing the bronchial tree from CT scans, but automatic bronchus classification is notoriously difficult because anatomical topology varies widely between patients. To tackle this, the researchers propose BCNet (Bronchus Classification Network), a structure-guided framework that leverages segment-level topological information from point clouds to enrich voxel-level representation learning. BCNet uses two jointly trained branches: a Point-Voxel Graph Neural Network (PV-GNN) that classifies bronchial segments, and a CNN that performs voxel-wise labeling. Both branches share a common convolutional backbone, enabling the topology-aware supervision from the PV-GNN to improve voxel-level features.

The key innovation is that during inference, only the CNN branch is required, meaning BCNet maintains the computational efficiency of its CNN baseline while benefiting from rich graph-based structure learning during training. Experiments on BronAtlas, a new open-access benchmark introduced in this paper, show BCNet outperforms state-of-the-art methods by more than 8.0% in F1-score for bronchus classification. BronAtlas features high-quality voxel-wise annotations of both anatomical and abnormal bronchial segments, providing a valuable resource for developing and evaluating advanced methods in bronchial tree analysis, disease diagnosis, and surgical planning. With the code and dataset publicly available, this work offers a practical pathway toward more reliable and efficient respiratory imaging AI.