SUMMARY

Affecting nearly 3% of adults above 65 years of age, aortic stenosis is a disease that involves narrowing of the aortic valve opening, which restricts the blood flow from the heart. If severe, the stenotic valve is typically replaced with a bioprosthetic valve via either an open-heart surgery or a transcatheter aortic valve replacement (TAVR), a less invasive procedure recommended for patients at increased risk of mortality with conventional surgery. A TAVR requires a careful procedural planning to consider the patients' aortic root anatomy and avoid any adverse effects that may be life-threatening. Thus, an accurate visualization and measurement of the aortic structures is paramount in the procedural planning of TAVR. In order to accurately visualize the anatomical structures of the aortic root and understand the fluid dynamics of the blood flow through the valve, a 3D reconstruction of the root geometry can be formulated by segmenting the anatomy from computer tomography (CT) images. Yet, a manual segmentation of the aortic root is a process both demanding and time-consuming. Instead, an efficient and accurate modeling of the aortic root geometry can be achieved using a set of unique aortic landmarks that can be localized on CT images. Accordingly, this research aims to utilize image processing and artificial intelligence-based methods such as convolutional neural networks to 1) automatically detect and localize the aortic landmarks of interest from CT images and 2) construct a 3D patient-specific reconstruction of the aortic root using the detected aortic landmarks. This study presents a method that can be utilized in assisting the pre-procedural planning of TAVR through fast, automatic segmentation of the aortic root structures.