Interpretable Diagnosis of Pulmonary Emphysema on Low-Dose CT Using ResNet Embeddings.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Accurate detection of pulmonary emphysema on low-dose computed tomography (LDCT) is crucial for screening and population studies.
• Existing methods face challenges in interpretability and scalability.

Purpose of the Study:

• To develop a quality-controlled and interpretable deep learning pipeline for emphysema assessment using LDCT.
• To enhance diagnostic performance and robustness by integrating deep learning embeddings with quantitative CT markers.

Main Methods:

• Utilized a ResNet-152 deep learning model for feature extraction from LDCT mid-lung patches.
• Implemented automated lung segmentation, quality-control filtering, and feature selection (logistic regression, LASSO, RFE).
• Fused ResNet embeddings with quantitative CT (QCT) markers (%LAA, Perc15, TLV) for improved analysis.

Main Results:

• Achieved high diagnostic performance with ROC-AUC of 0.996 and PR-AUC of 0.962.
• Demonstrated a balanced accuracy of 0.931 with low computational cost.
• Validated the effectiveness of pre-trained ResNet embeddings without retraining for emphysema characterization.

Conclusions:

• The proposed pipeline offers a reproducible and explainable framework for emphysema detection in population-level LDCT analysis.
• This approach supports research and serves as a screening-support tool for pulmonary emphysema.
• Deep learning embeddings, when combined with QCT markers, significantly enhance the robustness and interpretability of emphysema assessment.