A fully automated CT-based airway segmentation and branch labeling algorithm using deep learning and conventional image processing

Chronic obstructive pulmonary disease (COPD) is a chronic inflammatory lung disease that causes obstructed airflow from the lungs. COPD is the fourth leading cause of death in the United States and currently affects 328 million people worldwide with a projected increase in healthcare costs from $32 billion in 2010 to $49 billion in 2020. Quantitative computed tomography (CT)-based characterizations of bronchial metrics, such as airway lumen area and wall thickness, and parenchymal characterizations of emphysema and air trapping, are garnering research interest to help understand the pathophysiology and mechanism of several lung diseases' occurrence and progression. Airway segmentation and anatomical branch labeling allow spatial matching and referencing across individuals in large multi-center population-based studies such as, COPDGene, SPIROMICS, CanCOLD, MESA Lung, and SARP and has led to novel insights between COPD and airway phenotypes.

In this thesis, new theory and algorithms to automatically segment the human airway tree using thoracic CT imaging and label segmental bronchi are developed and evaluated. First, an image processing framework, called freeze-and-grow, is developed for airway tree segmentation which uses a multi-parametric approach to iteratively capture finer details of the airway tree while detection and correction for segmentation leakages. Then, deep learning is used to enhance the airway lumen in chest CT images to address inherent challenges in CT intensity-based airway tree segmentation while improving efficiency. The new airway segmentation methods outperform an industry standard method requiring manual intervention. Lastly, an automated algorithm for anatomical airway branches is presented which uses two stage machine learning and hierarchical features to compartmentalize the labeling process based on the airway tree branching anatomy.