PhD defence A. (Antonio) Garcia-Uceda Juarez

Dissertation in short:

In this thesis we developed automatic image processing methods to segment the bronchial tree from chest CT scans and subsequently extracted airway measurements. First, we developed a fully automatic airway segmentation method based on state-of-the-art deep learning algorithms. This method can obtain high-quality and complete airway segmentations on CT scans of different characteristics and including severe airway abnormalities. We built software around this method and made it publicly available, so that it can be used for future research. Next, we proposed two extensions to this method in order to improve its sensitivity, particularly the number of segmented peripheral airways. These are 1) a novel method that corrects typical segmentation errors made by our deep learning method, using synthetic labels augmented with errors, and 2) a combination of our deep learning method with graph neural networks, which are efficient to process tree-like structures such as the airway tree. Next, we complemented our deep learning method with two useful tools, namely 1) a method to generate ground truth airway segmentations needed for training of deep learning, and 2) a method to automatically segment both the lumen and wall surfaces of the bronchial tree. Finally, we studied the application of automated airway biomarkers to quantitatively assess structural airway abnormalities from CT scans in patients with lung diseases. These biomarkers are the airway tapering, airway-artery ratio and wall-artery ratio. With these developed methods, we aim to contribute to the development of automated tools for clinicians to assess accurately and efficiently pulmonary diseases affecting the airways.