Miranda Kirby

Ph.D

Toronto Metropolitan University

Assistant Professor

NLST (Learn more about this study)

NLST-1019

Feb 28, 2023

Rad-Ens: Texture-based Radiomics Machine Learning and Deep Learning Ensemble for COPD Classification using CT Images

Chronic obstructive pulmonary disease (COPD) is a progressive lung disease that is characterized by airflow limitation. The goal of COPD management is early detection to allow early treatment initiation to slow the disease progression. Computed tomography (CT) images can be obtained to detect abnormalities in the lung and to monitor COPD. Although COPD is diagnosed by spirometry, which measures lung function, it is unable to detect early structural changes and is not always acquired with lung CT images, for example in lung cancer screening trials and radiation treatment planning. To utilize CT images, machine learning (ML) and deep learning (DL) models have been used to classify COPD. However, COPD classification is more challenging in mild COPD with more subtle disease changes in the lung. We propose an ensemble of a texture-based radiomics ML and DL model, Rad-Ensbinary, for classifying binary COPD and Rad-Ensmulti for classifying COPD severity. These models will be tested using participants from Canadian Cohort Obstructive Lung Disease (CanCOLD), a multicenter population-based study of mainly mild COPD participants. For the machine learning model, 95 texture-based radiomics features will be extracted from the CT images and used with a feature selection method to select the top 5 features and a machine learning classifier for the predictions. For the deep learning model, a 3D VGG model will be utilized with 15 slices per subject. To create the RadEns models, the machine learning and deep learning outcome probabilities will be combined. Then to compare the performance of the machine learning, deep learning and Rad-Ens models, DeLong's test will be used. These Rad-Ens models may have the potential to be a computer-aided diagnosis (CAD) application that can be implemented when CT images are available. This would allow for COPD detection in those with undiagnosed COPD that undergo CT images and do not acquire spirometry.