Structure and position-aware graph neural network for airway labeling.

Area of Science:

• Medical Imaging
• Graph Neural Networks
• Computational Anatomy

Background:

• Accurate anatomical labeling of airway trees is crucial for diagnosing and managing respiratory diseases like Chronic Obstructive Pulmonary Disease (COPD).
• Existing methods for airway labeling often struggle with complex tree structures and variations in disease severity.

Purpose of the Study:

• To develop and evaluate a novel graph-based approach for automated anatomical labeling of airway tree segments.
• To improve the accuracy and efficiency of airway labeling compared to existing methods.

Main Methods:

• A graph-based method was developed, treating airway labeling as a branch classification problem.
• Features were extracted using convolutional neural networks (CNNs) and enhanced with structure-aware and position-aware graph neural networks (GNNs).
• The method was evaluated on 220 airway trees from COPD patients.

Main Results:

• The proposed method achieved an average accuracy of 91.18% for labeling 18 segmental airway branches, outperforming standard CNN (83.83%) and existing methods (87.37%).
• The approach demonstrated computational efficiency.
• A reader study indicated performance comparable to human experts on a separate set of 40 subjects.

Conclusions:

• The novel graph-based approach significantly enhances the accuracy of airway anatomical labeling.
• This method offers a computationally efficient and highly accurate tool for analyzing airway trees, particularly in the context of COPD.
• The algorithm's performance rivals that of human experts, suggesting its potential for clinical application.