A self-supervised framework for emphysema anomaly detection and staging in computed tomography scans.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Emphysema is a significant phenotype of chronic obstructive pulmonary disease (COPD), contributing to morbidity and increasing lung cancer risk.
• Computed tomography (CT) is crucial for emphysema detection, but current deep learning models require extensive annotated data.
• Unsupervised anomaly detection (UAD) methods struggle with emphysema's specific anomalies and semantic representation.

Purpose of the Study:

• To develop a self-supervised learning framework for emphysema detection and staging using only non-emphysema CT scans.
• To address the limitations of existing UAD methods in identifying and localizing emphysema with weak semantics.
• To create a scalable and interpretable AI solution for lung disease analysis.

Main Methods:

• A self-supervised framework trained on non-emphysema CT scans with synthetically generated emphysema lesions.
• Introduction of EDLNet, an encoder-decoder architecture featuring spatial-channel refinement and adaptive feature fusion.
• Unsupervised emphysema staging integrated into the detection and localization framework.

Main Results:

• The proposed framework demonstrated superior performance in emphysema detection and localization compared to existing UAD approaches.
• Achieved a mean staging accuracy of 93.13% for emphysema.
• Attained a macro area under the receiver operating characteristic curve (AUROC) of 99.08% in multi-center evaluations.

Conclusions:

• The developed self-supervised framework offers a robust and efficient method for emphysema analysis without requiring large annotated datasets.
• EDLNet effectively bridges clinical knowledge with artificial intelligence for improved lung disease detection and staging.
• This approach provides a scalable, interpretable, and high-performing solution for analyzing emphysema on CT scans.