Logo image
Back
B110-11 Early Prediction of Rapid Emphysema Progression in a Lung Cancer Screening Cohort Using Quantitative LDCT and an Ensemble-based Machine Learning Model
Abstract   Peer reviewed

B110-11 Early Prediction of Rapid Emphysema Progression in a Lung Cancer Screening Cohort Using Quantitative LDCT and an Ensemble-based Machine Learning Model

K Knoernschild, K E Schroeder, J J Kitzmann and J C Sieren
American journal of respiratory and critical care medicine, Vol.212(Supplement_1), pp.S1479-S1480
05/01/2026
DOI: 10.1093/ajrccm/aamag162.1946

View Online

Abstract

Rationale Lung cancer screening detects malignant nodules in only 2-5% of participants. Opportunistic evaluation of other lung diseases, such as emphysema, could extend the utility of CT screening. Identifying individuals at risk for rapid emphysema progression may inform smoking cessation strategies and closer pulmonary follow-up, even in the absence of lung cancer. We propose an ensemble-based machine learning approach leveraging whole-lung quantitative radiomics features for opportunistic risk prediction at the time of lung cancer screening. Methods A cohort of 576 LDCT subjects with at least two timepoints were selected from the National Lung Screening Trial (NLST). Low Attenuation Areas less than -950 Hounsfield Units (LAA-950) were calculated for each timepoint. Rapid emphysema progressors were defined as having a change in LAA-950 ≥1% per year. Example progressor and non-Progressor cases with similar %LAA-950 at the time of screening are shown in Figure 1. The final cohort was split into Development (Dev.), Ensemble Validation (EnsVal.), and External (Ext.) datasets. Whole lung segmentations were created for the initial LDCT of each subject, followed by radiomic feature extraction, including first order, texture, shape, and wavelet feature variations. Radiomics features were harmonized to the GE:STANDARD reconstruction kernel cases in the Dev. dataset. Important features were selected from the Dev. dataset using Pearson’s correlation and a round-based ElasticNet logistic regression model, which were then used to train 1000 artificial neural networks (ANNs). The final ensemble consisted of the 100 best performing networks, chosen according to the EnsVal. dataset sensitivity and specificity metrics at the Youden index threshold. Final ensemble predictions were made, and performances were compared using AUC-ROC, sensitivity, and specificity for the Dev., EnsVal., and Ext. datasets. Results 36 important whole-lung radiomics features were chosen, with a majority being wavelets (30 features). All datasets achieved ≥0.71 AUC-ROC (Dev.: 0.91, EnsSel.: 0.75, Ext: 0.71). Sensitivity for progressors ranged from 0.77 (EnsSel.) to 0.95 (Dev.). Specificity values ranged from 0.56 (Ext.) to 0.70 (Dev.). Conclusions Current results show promise for an ensemble-based ANN approach for predicting emphysema progression using LDCT images at the time of lung cancer screening, with all datasets achieved a greater than 0.71 AUC-ROC. Planned improvements include standardizing slice-spacing via resampling prior to radiomic feature extraction and selection, followed by optimization of ensemble training and selection to increase sensitivity and specificity for clinical detection of rapid emphysema progressors. This abstract is funded by: National Institute of Health (R01CA267820)
 Lung Cancer Machine Learning Medical Screening Datasets Emphysema Lung diseases Radiomics

Details

Metrics

1 Record Views
Logo image