CT-Derived Functional Imaging for Predicting Disease Progression in COPD

PROJECT SUMMARY Chronic obstructive pulmonary disease (COPD) is the third leading cause of death in the United States, imposing a significant economic burden due to its high morbidity and mortality. The Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria is a universally accepted disease severity staging score. However, GOLD score is not a strong predictor for mortality or progression risk at early or pre-disease stages. Early intervention is crucial for slowing COPD progression and improving quality of life. Therefore, there is a need to develop robust, quantitative metrics for characterizing disease state and progression risk. Existing quantitative computer tomography (CT) methods are based on analyzing variations in CT Hounsfield Units (HU) and have shown moderate to strong correlation with disease state. However, HU values are known to be breathing-effort dependent. As a result, quantitative CT methods require heuristic normalization schemes to adjust for varying lung inflation levels and are known to lack reproducibility. We previously developed a robust class of CT-derived ventilation (CTV) methods that calculate breathing-induced volume changes apparent on inhale/exhale CT (IE- CT) image pairs, as a surrogate for ventilation. In addition to numerical stability, our CTV demonstrated a higher correlation with nuclear medicine-based ventilation imaging than any other method in the literature. We recently extended the CTV framework to calculate changes in blood mass apparent on IE-CT, as a surrogate for pulmonary perfusion. Our novel CT-Perfusion (CTP), taken together with CTV, comprise our CT-derived functional imaging (CTFI) methodology. CTFIis the first to mathematically describe changes in inhale/exhale HU values in terms of ventilation and perfusion. This allows us to compute VQ (ventilation/perfusion) ratio imaging that is inherently normalized to patient breathing effort. Thus, any early microvascular changes or VQ mismatch associated with COPD disease severity can potentially be detected and quantified on IE-CT images. We hypothesize that a CTFI-informed machine learning model has higher discriminative power in assessing survival and disease progression than traditional methods such as FEV1, BODE index and other quantitative imaging markers. To test this hypothesis, we will utilize data from the Genetic Epidemiology of COPD (COPDgene) study, a multicenter observational study designed to identify genetic factors associated with COPD. We will adapt state- of-the art unsupervised deep learning methods and fully leverage the rich COPDgene data set to train a lung lobe segmentation model and automate the CTFI calculation pipeline. Next, we will develop and validate both physics-based and deep learning-based CTFI VQ scoring methods. Finally, we will develop a machine learning model which takes clinical information and patient CTFI lobar scores (ventilation, perfusion, & VQ) as input, and quantifies disease severity, predicts 5-year survival, and predicts future exacerbation and progression risk. All developed software and training data will be made publicly available to encourage further research in this field.