COPD-TransNet: A Swin Transformer Network with Quantitative Emphysema Feature Fusion for COPD Detection and Staging...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Lung cancer screening computed tomography (CT) scans offer a potential avenue for screening chronic obstructive pulmonary disease (COPD).
• Accurate detection and staging of COPD are crucial for timely intervention and management.
• Current methods may not fully leverage the information available in screening CT scans for comprehensive COPD assessment.

Purpose of the Study:

• To develop and validate a deep learning model, COPD-TransNet, for detecting and staging COPD using lung cancer screening CT scans.
• To assess the model's performance based on Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria.
• To integrate CT image analysis with emphysema features and lung volume parameters (LAV-950%) for enhanced COPD classification.

Main Methods:

• A Swin Transformer-based deep learning architecture was employed for COPD detection, staging, and severity classification.
• The model integrated preprocessed CT images, emphysema features, and LAV-950% metrics.
• Training and testing involved a dataset of 637 patients from a pulmonary nodule clinic, with external validation using 1464 CT scans from the National Lung Screening Trial (NLST) cohort.

Main Results:

• COPD-TransNet achieved an AUC of 0.829 for COPD detection, outperforming mainstream methods.
• The model demonstrated strong performance in severity classification (F1 score 0.763, accuracy 0.791) and staging (accuracy 0.789).
• External validation on the NLST cohort yielded an AUC of 0.867, confirming the model's robustness and clinical feasibility.

Conclusions:

• The proposed COPD-TransNet framework effectively utilizes Swin Transformer architecture and LAV-950% features for COPD screening and staging.
• This deep learning approach shows significant potential for improving COPD detection and classification within lung cancer screening programs.
• The model's validated performance highlights its clinical feasibility for widespread application in COPD management.