Seminars

Abstract

Inspiratory-expiratory CT image pairs can be used to characterize the normal lung and identify lung disorders such as COPD. Emphysematous regions can be identified on the inspiratory scan, air trapping regions can be detected on the expiratory scan, and by registering the inspiratory and expiratory images, functional small airways disease (fSAD), local lung volume change, and biomechanical parameters can be computed. However, acquiring multiple CT images increases time, cost, and radiation dose. We describe a deep-learning approach that can directly estimate multi-volume, lung biomarkers from a single inspiratory CT image. Our approach uses a style-based generative adversarial network called Lung2Lung for translating CT images from end-inspiratory to end-expiratory volume. The input to the network is a CT image at inspiration and the network synthesizes the corresponding expiratory image. The network can synthesize an artificial expiratory image from a real inspiratory image in about 40 seconds. The expiratory image is created in the shape of the inspiratory image, so effectively, the two images are in registration and local lung volume change and parametric response mapping biomarkers can be directly computed. This approach can enable air trapping, fSAD, and lung volume change analysis in existing data sets that only contain inspiratory images, and reduce time, cost, and radiation dose by allowing new studies to collect only inspiratory scans. It may be possible to extend this approach to other related image translation tasks, such as estimating a synthetic hyperpolarized gas MRI image directly from a CT image or proton MRI image.