Emphysema subtyping on thoracic computed tomography scans using deep neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Accurate emphysema subtyping and severity assessment are vital for managing Chronic Obstructive Pulmonary Disease (COPD) and understanding disease heterogeneity.
• Manual analysis is time-consuming and subjective, limiting clinical utility and research.
• Existing automated methods have shown moderate performance.

Purpose of the Study:

• To develop and validate a deep learning-based automated system for emphysema subtyping and severity analysis.
• To replicate the Fleischner Society's visual scoring system using artificial intelligence.
• To improve upon existing automated methods for emphysema assessment.

Main Methods:

• A deep learning algorithm was trained and evaluated on 9650 subjects from the COPDGene study.
• The approach utilized a regression training strategy for categorical label generation.
• High-resolution localized activation maps were generated for visualization and quantitative analysis.

Main Results:

• The deep learning algorithm achieved a predictive accuracy of 52%, outperforming a previous method's accuracy of 45%.
• The agreement between predicted and visual scores was good, surpassing the moderate agreement of the previous method.
• The method can compute emphysema involvement percentage per lung and identify both centrilobular and paraseptal emphysema subtypes.

Conclusions:

• The proposed deep learning approach offers an accurate and automated solution for emphysema subtyping and severity analysis.
• This method provides enhanced predictive accuracy and agreement compared to prior techniques.
• The system's ability to generate activation maps enables detailed quantitative assessment and extends subtype identification.