Machine Learning and CT Texture Features in Ex-smokers with no CT Evidence of Emphysema and Mildly Abnormal Diffusing...

Ex-smokers without spirometry or CT evidence of chronic obstructive pulmonary disease (COPD) but with mildly abnormal diffusing capacity of the lungs for carbon monoxide (DL_CO) are at higher risk of developing COPD. It remains difficult to make clinical management decisions for such ex-smokers without other objective assessments consistent with COPD. Hence, our objective was to develop a machine-learning and CT texture-analysis pipeline to dichotomize ex-smokers with normal and abnormal DL_CO (DL_CO≥75%_pred and DL_CO<75%_pred).

In this retrospective study, 71 ex-smokers (50-85yrs) without COPD underwent spirometry, plethysmography, thoracic CT, and ³He MRI to generate ventilation defect percent (VDP) and apparent diffusion coefficients (ADC). PyRadiomics was utilized to extract 496 CT texture-features; Boruta and principal component analysis were used for feature selection and various models were investigated for classification. Machine-learning classifiers were evaluated using area under the receiver operator characteristic curve (AUC), sensitivity, specificity, and F1-measure.

Of 71 ex-smokers without COPD, 29 with mildly abnormal DL_CO had significantly different MRI ADC (p < .001), residual-volume to total-lung-capacity ratio (p = .003), St. George's Respiratory Questionnaire (p = .029), and six-minute-walk distance (6MWD) (p < .001), but similar relative area of the lung < -950 Hounsfield-units (RA₉₅₀) (p = .9) compared to 42 ex-smokers with normal DL_CO. Logistic-regression machine-learning mixed-model trained on selected texture-features achieved the best classification accuracy of 87%. All clinical and imaging measurements were outperformed by high-high-pass filter high-gray-level-run-emphasis texture-feature (AUC=0.81), which correlated with DL_CO (ρ = -0.29, p = .02), MRI ADC (ρ = 0.23, p = .048), and 6MWD (ρ = -0.25, p = .02).

In ex-smokers with no CT evidence of emphysema, machine-learning models exclusively trained on CT texture-features accurately classified ex-smokers with abnormal diffusing capacity, outperforming conventional quantitative CT measurements.