A self-training framework for semi-supervised pulmonary vessel segmentation and its application in COPD.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Accurate segmentation of pulmonary vessels in CT images is vital for diagnosing and managing chronic obstructive pulmonary disease (COPD).
• Smaller vessels are particularly challenging to segment accurately.
• Existing methods may struggle with the complexity of pulmonary vasculature in COPD patients.

Purpose of the Study:

• To develop and evaluate a semi-supervised method for segmenting pulmonary vasculature in CT images.
• To improve the precision of pulmonary vessel segmentation, especially in COPD patients.
• To enable quantitative analysis of pulmonary vessels in relation to COPD severity.

Main Methods:

• A self-training framework utilizing a teacher-student model for semi-supervised pulmonary vessel segmentation.
• Interactive annotation for high-quality initial labels.
• Iterative training using pseudo-labels generated by a teacher model and selected based on reliability.
• Experiments conducted on non-enhanced CT scans from 125 COPD patients.

Main Results:

• The proposed Semi2 method achieved a precision of 90.3% for vessel segmentation, a 2.3% improvement.
• Quantitative analysis provided insights into pulmonary vessel differences across varying COPD severity.
• Both quantitative and qualitative analyses confirmed the method's effectiveness.

Conclusions:

• The Semi2 method significantly enhances pulmonary vascular segmentation performance.
• The approach is applicable for detailed COPD analysis, offering insights into disease-specific vascular changes.
• The developed code will be publicly available for further research.