Adversarial Transformer for Repairing Human Airway Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Automated airway segmentation models struggle with peripheral bronchioles, limiting clinical use.
• Data heterogeneity and lung pathologies challenge accurate segmentation of small airways.

Purpose of the Study:

• To develop a robust airway segmentation method addressing discontinuities and improving accuracy in diverse lung conditions.
• To enhance the clinical applicability of automated airway segmentation for lung disease diagnosis and prognosis.

Main Methods:

• A patch-scale adversarial-based refinement network was proposed, utilizing preliminary segmentation and original CT images.
• The method was validated on datasets including healthy, pulmonary fibrosis, and COVID-19 cases.
• Quantitative evaluation used seven metrics, and visual results were analyzed for discontinuity detection.

Main Results:

• The proposed method achieved over a 15% increase in detected length and branch ratios compared to existing models.
• The refinement approach effectively detected discontinuities and missing bronchioles, improving segmentation completeness.
• The pipeline generalized well, significantly enhancing the performance of three previous segmentation models.

Conclusions:

• The developed method offers a robust and accurate tool for airway segmentation in medical imaging.
• Improved segmentation accuracy aids in better diagnosis and treatment planning for various lung diseases.
• The approach shows promise in overcoming limitations of current automated segmentation techniques.