Muhammad Faizyab Chaudhary

Abstract: Chronic obstructive pulmonary disease (COPD) is a complex disorder that leads to progressive decline in lung function and poor quality of life. Over the past decade, a number of interesting developments in quantitative computed tomography (QCT) imaging have enabled the early characterization of small airways disease, a well-known precursor to COPD. Concurrently, image registration of multiple CT volumes has enabled local characterization of biomechanical defects across the lungs. These spatial measures of functional abnormalities have been associated with several outcomes in COPD, including overall lung function, symptom burden, and the risk of exacerbations. The strength of these measures stems from their ability to localize disease phenotypes, thereby opening possible avenues into mechanisms. These developments have thus introduced the notion of locality for understanding lung disease. Although several large multicenter studies have carefully acquired computed tomography scans at different volumes for a better functional assessment of COPD, this practice remains far from being protocolled in the clinic. Typically, single CT volumes at inspiration are acquired in most clinical settings, which precludes a large bulk of clinical data from being analyzed for functional characterization. In addition, several retrospective multicenter studies were not designed to acquire multiple CT images. Our work attempts to solve this problem by introducing deep generative modeling as a new paradigm for deriving functional surrogates of local lung function from single-volume CT scans. We hypothesized that a CT scan acquired at a single volume can be used to derive different measures of local lung function by learning a deep generative adversarial image synthesis model. We propose a novel high-resolution medical image-to-image translation framework for cross-volume computed tomography image-to-image translation. We demonstrated that our models could be used to estimate different measures of local lung function, including the parametric response mapping-based functional small airways disease, image registration-based local tissue expansion, and the hyperpolarized gas magnetic resonance images. We validated our deep generative models in two large ongoing cohort studies of COPD – SPIROMICS and COPDGene.