Radiomics Analysis Improves Machine Learning Models Assessing Chronic Obstructive Pulmonary Disease
R Au1*, K Makimoto1, W Tan2, J Bourbeau3, J Hogg2, M Kirby1, (1) Ryerson University, Toronto, ON, CA, (2) Centre For Heart Lung Innovation, University Of British Columbia, Vancouver, BC, CA, (3) McGill University, Montreal, QC, CA
Presentations
PO-GePV-M-40 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Computed tomography (CT) imaging can be used to quantitatively investigate chronic obstructive pulmonary disease (COPD). However, existing quantitative CT (QCT) features only assess structural characteristics (e.g. volumes and counts). Texture-based radiomics analysis is an emerging quantitative technique that investigates spatial relationships between voxel intensities and may therefore provide increased sensitivity to early or subtle disease changes. We hypothesized that CT radiomics features, in combination with existing QCT features, will increase accuracy for classifying COPD.
Methods: Spirometry and CT images were obtained from the population-based CanCOLD study. Well-established QCT features consisted of the percentage of low-attenuation-areas below -950HU and -856HU, 15th percentile of the CT-density-histogram, low-attenuation-cluster slope, CT-tissue-volume, total-airway-count, and hypothetical airway of 10mm lumen perimeter (VIDA Diagnostics). For radiomics analysis, CT images were processed using whole lung and airway segmentation, voxel resampling, and thresholding between -1000HU-to-0HU prior to extraction of grey-level-co-occurrence, grey-level-run-length, grey-level-size-zone, grey-level-distance-zone, neighborhood-grey-tone-difference, and neighboring-grey-level-dependence texture-based radiomics features. Highly correlated features were removed and a least absolute shrinkage and selection operator feature selection method was completed. 5 iterations of a nested 10-fold cross validation training was performed using a Gaussian support vector machine with hyperparameter optimization. The mean classification accuracy was used to assess performance. Multivariable linear regression models for forced-expiratory-volume-in-1-second/forced-vital-capacity (FEV₁/FVC) were constructed.
Results: A total of 1174 participants were evaluated (No COPD: n=582; COPD: n=592). The model trained with radiomics and QCT features had a 74.4% accuracy in classifying COPD, and was significantly greater than the accuracy from the only QCT features trained model (72.5%) (p<0.0001). The combined feature model also had a stronger correlation with FEV₁/FVC (r²=0.53, p<0.0001) than the only QCT features model (r²=0.47, p<0.0001).
Conclusion: CT radiomics features provides additional and complementary information when used with QCT features for characterizing COPD. Future studies should investigate using radiomics analysis for longitudinal COPD assessment.
Funding Support, Disclosures, and Conflict of Interest: This research was funded by NSERC. Dr. Kirby gratefully acknowledges support from the Parker B. Francis Fellowship Program and the Canada Research Chair Program (Tier II). Dr. Kirby is a consultant for Vida Diagnostics Inc.
ePosters
Keywords
Lung, Quantitative Imaging, CT
Taxonomy
Contact Email