MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

CT Image Parameter Estimation Using PCA-Based Deep Learning in Chronic Obstructive Pulmonary Disease

A Moslemi1*, W Tan2, J Bourbeau3, J Hogg2, H Coxson2, M Kirby1, (1) Ryerson University, Toronto, Ontario, CA, (2) Centre For Heart Lung Innovation, University Of British Columbia, (3) McGill University

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide. Smokers with normal lung function have evidence of abnormalities on computed tomography (CT). Although CT is not part of COPD standard-of-care, predicting at risk individuals with structural abnormalities may enable earlier detection. Our objective is to use principle component analysis (PCA)-based deep learning to estimate CT measurements in at risk smokers using only clinical measurements.

Method: CanCOLD study participants were evaluated using 6 clinical (age, BMI, pack years, dyspnea, chronic productive cough, wheezing) and seven pulmonary function test measurements (FEV1, FVC, FEV¬1/FVC, FEF25-75, TLC, RV and FRC). CT measurements included airway wall thickness (Pi10) and emphysema (low-attenuation-area-below-950HU: LAA950) (VIDA Diagnostics). A total of n=266 never-smokers, n=115 at risk, n=349 mild COPD, n=266 moderate COPD were used for training; a total of n=250 at risk participants were used for testing. PCA method using feature selection followed by Z-score normalization was used. The deep neural network was trained using 3-fully connected layers with “Swish” activation for hidden layers. K-fold cross validation and 20% drop out were used to tune number of epochs and neural network topology, respectively. Mean absolute percentage error (MAPE) is the percentage of the deviation of the predicted from true values.

Results: The features selected by PCA were FEV1, TLC, RV, FEF25-75 and FRC. The estimated/actual LAA950 and Pi10 measurements were 69.5±9.0%/69.8±9.3% and 4.00±0.15mm/3.97±0.15mm, respectively. The MAPE for LAA950 was 1.1% and for Pi-10 was 3.7%. The r-square for the goodness of fit of model was 0.51 and 0.84 for Pi-10 and LAA950, respectively.

Conclusion: CT parameters can be estimated using only clinical and spirometry measurements. Estimation of structural abnormalities may allow at risk individuals to be stratified to receive CT imaging to confirm abnormalities, and enable more careful follow-up/management to reduce adverse health-related outcomes.

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.

Keywords

CT, Lung, Quantitative Imaging

Taxonomy

IM- CT: Quantitative imaging/analysis

Contact Email