Fully-Automated Pulmonary Hyperpolarized Magnetic Resonance Image Registration, Segmentation and Quantification of Ventilation Defects Using Convolutional Neural Networks

Asthma and chronic obstructive pulmonary disease (COPD) are heterogeneous diseases characterized by limited or absent airflow due to airway obstruction, tissue destruction, and airway remodeling. Clinically, pulmonary function tests, such as spirometry and forced expiratory volume in 1 second (FEV1), are commonly used for patient diagnosis and monitoring. FEV1 provides a whole-lung measurement of airflow obstruction. It is not sensitive to detecting the underlying disease pathophysiology and its relevance to regional obstruction information, limiting its ability to guide therapy. This limitation has driven the development of medical imaging approaches sensitive to pulmonary abnormalities, including inhaled hyperpolarized noble gas magnetic resonance imaging (MRI), which uses helium-3 (3He) and xenon-129 (129Xe) gases. Hyperpolarized gas MRI enables direct visualization of ventilation, with heterogeneity due to reduced or absent airflow appearing as regions with minimal or absent signal. These regions are more appropriately labeled as ventilation defects and are directly quantified using ventilation defect percent (VDP). VDP is computed as the ratio of non-ventilated lung volume from 129Xe MRI to the thoracic cavity volume from proton (1H) MRI. While early methods relied on subjective scoring, manual segmentation, image processing, and semi-automated methods, contemporary approaches utilize deep learning, specifically convolutional neural networks (CNNs). However, most CNN-based VDP quantification methods lack external validation, limiting their application in multicenter studies. This thesis aims to develop and validate an automated deep-learning pipeline for 129Xe MRI VDP quantification that is generalizable and translatable to clinical and research settings.