Whole-Lung CT Radiomics-Based Machine Learning Classification of Nontuberculous Mycobacterial Lung Disease Across...

Area of Science:

• Radiology
• Artificial Intelligence
• Pulmonology

Background:

• Nontuberculous mycobacterial lung disease (NTM-LD) presents diagnostic challenges due to its heterogeneity.
• Previous CT radiomics studies for NTM-LD relied on manual segmentation, limiting scalability.
• This research introduces an automated, whole-lung CT radiomics approach for NTM-LD identification.

Purpose of the Study:

• To develop and validate a machine learning model using whole-lung CT radiomics for classifying NTM-LD.
• To assess the model's generalizability across geographically distinct patient cohorts.
• To identify common radiomic features indicative of NTM-LD.

Main Methods:

• Utilized 1,300 chest CT scans from China and 173 from the US (NTM cohort).
• Employed automated whole-lung segmentation and extracted 85 quantitative radiomic features.
• Evaluated model performance using two frameworks: single-dataset training with external validation and combined-cohort training, with Linear Discriminant Analysis (LDA) as the primary classifier.

Main Results:

• The model trained on the Chinese cohort achieved an AUC of 0.79, with external validation on the US cohort yielding an AUC of 0.94.
• Training on the combined cohort resulted in an AUC of 0.81, improving sensitivity and precision.
• Feature importance analysis highlighted 16 consistent texture-based features reflecting distinct parenchymal patterns.

Conclusions:

• Whole-lung CT radiomics provides an interpretable and automated method for NTM-LD classification.
• The findings suggest the existence of population-independent parenchymal signatures for NTM-LD.
• This approach facilitates timely and accurate identification of NTM-LD across diverse populations.