Hearing the unheard: single CT volume surrogates of regional lung function

Human lungs help us breathe around the clock, and their normal function is very tightly regulated by a concert of various mechanisms. This intricate harmony can be disrupted in countless different ways, and these disruptions can often lead to some very complicated lung disorders. One such complication of the lungs is the chronic obstructive pulmonary disease (COPD), which presents in different yet complicated ways. One of the emerging clinical tools for understanding and characterizing COPD is computed tomography (CT) imaging which allows us to visualize lung tissues and the airways under diseased conditions. Computed tomography imaging has helped the clinicians devise various interventions for managing COPD. Computational tools have also been used to automatically analyze CT scans of COPD patients and several clinically relevant outcomes have thus been developed for understanding disease stage and therapeutic response. One can easily wonder: are there any well-known CT biomarkers? if there are any, have they been validated and tested in diverse clinical conditions? The answer to the first question is yes: CT scans acquired at full-exhale and full-inhale have been able to provide some clinically relevant measures of the overall and local lung function. The most interesting thing about these biomarkers, that helps them stand apart from other approaches, is their ability to localize lung regions impacted by COPD – locality. This comes with a drawback: these approaches require at least two CT scans, however, in most clinical settings CT scans are acquired at a single volume.

While the single volume CT data continues to grow very rapidly around the world, there are not local CT imaging measures that could help characterize lung function. We thought to ourselves, if the very recent generative models can be used to transform human faces to sketches or doodles, or if horses can be converted into zebras, why cannot we use generative modeling for synthesizing CT scans at one volume from another? This could be really useful, since we could then synthesize scans where only one CT scan was acquired and perform the functional analysis. This thesis is all about the development of such generative models. It takes you through the process of estimating functional lung imaging measures using single CT volumes using deep generative modeling.