Graph refinement based airway extraction using mean-field networks and graph neural networks

Area of Science:

• Medical imaging analysis
• Computer vision
• Graph theory

Background:

• Graph refinement is crucial for extracting meaningful subgraphs from complex data.
• Extracting tree structures from volumetric image data, such as airways in CT scans, presents a significant challenge.

Purpose of the Study:

• To develop and compare novel graph refinement methods for tree extraction from image data.
• To apply these methods for accurate airway segmentation in 3D chest CT scans.

Main Methods:

• Graph-based representation of volumetric data followed by graph refinement.
• Mean-field approximation (MFA) implemented as mean-field networks (MFNs) for subgraph estimation.
• Supervised learning using graph neural networks (GNNs) for direct edge probability prediction.

Main Results:

• Both MFN and GNN models significantly improved airway detection compared to baseline and 3D U-Net methods.
• The proposed methods identified more airway branches with a reduction in false positives.
• Demonstrated effectiveness on 3D, low-dose chest CT data.

Conclusions:

• Graph refinement offers a powerful framework for extracting complex structures like airways from medical images.
• MFNs and GNNs provide effective and comparable solutions for this task, outperforming existing segmentation models.
• These methods hold promise for enhanced analysis of medical imaging data.