TNN: Tree Neural Network for Airway Anatomical Labeling

Area of Science:

• Medical imaging analysis
• Graph neural networks
• Computational anatomy

Background:

• Accurate anatomical labeling of bronchial trees from CT images is crucial for intra-operative navigation.
• Challenges include sparse airway voxels, class imbalance, and overlapping features in 3D image data.

Purpose of the Study:

• To develop a graph-neural-network-based method for detailed anatomical labeling of bronchial trees.
• To address challenges in airway voxel distribution and feature overlap, especially for subsegmental labeling.
• To enable fine-grained maps for improved intra-operative navigation.

Main Methods:

• A graph neural network (GNN) approach maps bronchial branches to graph nodes for anatomical labeling.
• Focuses on relative positions of sibling subsegments to overcome feature overlap.
• Utilizes multi-level labeling for hierarchical nomenclature and hyperedges for subtree representation.
• A hypergraph neural network encodes subtree relationships, guided by a filter module for feature aggregation.

Main Results:

• Achieved 93.6% accuracy for segmental node classification.
• Achieved 82.0% accuracy for subsegmental node classification.
• The method effectively handles sparse data and class imbalance in 3D airway imaging.

Conclusions:

• The proposed GNN method provides accurate and detailed anatomical labeling of bronchial trees.
• This approach enhances the potential for fine-grained intra-operative navigation using CT-derived maps.
• The developed method offers a robust solution for complex airway labeling tasks.