Doctor of Philosophy
Dr. Grace Parraga
Dr. Aaron Fenster
Pulmonary imaging, including pulmonary magnetic resonance imaging (MRI) and computed tomography (CT), provides a way to sensitively and regionally measure spatially heterogeneous lung structural-functional abnormalities. These unique imaging biomarkers offer the potential for better understanding pulmonary disease mechanisms, monitoring disease progression and response to therapy, and developing novel treatments for improved patient care. To generate these regional lung structure-function measurements and enable broad clinical applications of quantitative pulmonary MRI and CT biomarkers, as a first step, accurate, reproducible and rapid lung segmentation and registration methods are required. In this regard, we first developed a 1H MRI lung segmentation algorithm that employs complementary hyperpolarized 3He MRI functional information for improved lung segmentation. The 1H-3He MRI joint segmentation algorithm was formulated as a coupled continuous min-cut model and solved through convex relaxation, for which a dual coupled continuous max-flow model was proposed and a max-flow-based efficient numerical solver was developed. Experimental results on a clinical dataset of 25 chronic obstructive pulmonary disease (COPD) patients ranging in disease severity demonstrated that the algorithm provided rapid lung segmentation with high accuracy, reproducibility and diminished user interaction. We then developed a general 1H MRI left-right lung segmentation approach by exploring the left-to-right lung volume proportion prior. The challenging volume proportion-constrained multi-region segmentation problem was approximated through convex relaxation and equivalently represented by a max-flow model with bounded flow conservation conditions. This gave rise to a multiplier-based high performance numerical implementation based on convex optimization theories. In 20 patients with mild- to-moderate and severe asthma, the approach demonstrated high agreement with manual segmentation, excellent reproducibility and computational efficiency. Finally, we developed a CT-3He MRI deformable registration approach that coupled the complementary CT-1H MRI registration. The joint registration problem was solved by exploring optical-flow techniques, primal-dual analyses and convex optimization theories. In a diverse group of patients with asthma and COPD, the registration approach demonstrated lower target registration error than single registration and provided fast regional lung structure-function measurements that were strongly correlated with a reference method. Collectively, these lung segmentation and registration algorithms demonstrated accuracy, reproducibility and workflow efficiency that all may be clinically-acceptable. All of this is consistent with the need for broad and large-scale clinical applications of pulmonary MRI and CT.
Guo, Fumin, "Pulmonary Image Segmentation and Registration Algorithms: Towards Regional Evaluation of Obstructive Lung Disease" (2017). Electronic Thesis and Dissertation Repository. 4746.