MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning and Texture Analysis of Thoracic X-Ray Computed Tomography to Reveal Subclinical Emphysema

M Sharma1,2*, AR Westcott1,2, JL MacNeil1,3, DG McCormack4 and G Parraga1-4, (1) Robarts Research Institute, (2) Department of Medical Biophysics, (3) School of Biomedical Engineering, (4) Department of Medicine, Western University, London, ON, Canada

Presentations

(Sunday, 7/12/2020) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose: Diffusion-weighted hyperpolarized gas magnetic resonance imaging (MRI) pulmonary measurements are highly sensitive to mild airspace enlargement (or subclinical emphysema) in ex-smokers in whom pulmonary CT does not reveal any parenchymal abnormalities. Our aim was to use machine-learning and texture analysis to uncover the novel CT features that explain MRI apparent diffusion coefficients consistent with airspace enlargement or emphysema.


Methods: Ex-smokers without spirometry evidence of airflow limitation provided informed consent to an approved protocol and underwent spirometry, hyperpolarized gas MRI and thoracic CT. Participants were dichotomized based on diffusing capacity of the lungs for carbon-monoxide (DLCO) threshold of =75%. CT texture features were extracted from a novel 3-dimensional extension of gray-level run-length-matrices, gap-length-matrices, zone-size-matrices and co-occurrence-matrices. Principal component analysis and forward selection logistic-regression were used to select CT features that significantly contributed to model accuracy. Classification models included support-vector-machines, decision-trees, logistic-regression and nearest-neighbour classifiers. Model performance was evaluated using area-under-the-receiver-operator-characteristic-curve (AUC), sensitivity and specificity.


Results: We evaluated 65 ex-smokers (Normal DLCO: n=41, Abnormal DLCO: n=24) without significant differences in pulmonary function tests and relative-area-of-the-lung <-950 Hounsfield Units (RA950) between groups. Long-run-emphasis (LRE) and run-percentage (RP) were statistically different (p=0.001 and p=0.022) between groups and significantly but moderately correlated with MRI apparent diffusion coefficient measurements (r=0.48, p<0.001 and r=-0.46, p<0.001 respectively). These features strongly contributed to all models while the Quadratic-SVM achieved a modest classification accuracy of 80%.


Conclusion: Using machine-learning we identified CT texture features that differentiated ex-smokers with and without mild emphysema not revealed using conventional CT measurements such as RA950. In other words, machine-learning revealed that greater long run emphasis and diminished run percentage reflect homogeneous CT textures that are related to mild parenchymal emphysematous disease. This is important because until now it has been impossible to measure mild airspace enlargement using only CT lung tissue attenuation values.

Keywords

Not Applicable / None Entered.

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

Contact Email