PROJECT SUMMARY/ABSTRACT Transcatheter Aortic Valve Replacement (TAVR) is an emerging treatment option for aortic stenosis, a common heart valve disease that causes about 15,000 deaths per year in the U.S. TAVR has been steadily gaining popularity since 2011, and is now performed over 70,000 times per year in the U.S. Finite element (FE) methods have shown great potential for improving TAVR treatment planning by simulating the biomechanical interactions between anatomical structures and deployed prosthetic devices. However, FE methods are currently severely limited by the delineation process of patient-specific geometry, as manual delineation from 3D CT images is extremely time consuming and error-prone. Automated methods have been proposed, but they have limited adaptability due to extensive assumptions about input and output characteristics. This is especially problematic when extensions of patient-specific geometry are required to simulate various complications of TAVR. To address these limitations, this proposal aims to develop fast, robust, and easily adaptable deep learning algorithms for automating the delineation of patient-specific geometry from 3D CT images. Aim 1 is to develop template deformation-based weakly supervised deep learning algorithms to delineate TAVR-relevant anatomical structures such as the upper left ventricular myocardium, aortic valve, coronary arteries, and ascending aorta. The template deformation strategy will establish mesh correspondence between all predicted volumetric FE outputs, and weak supervision will allow for modeling of the complex output geometry with minimally sufficient expert labeling. Aim 2 is to incorporate anatomically consistent calcification to the final mesh outputs using multi-task deep learning. Based on prior medical knowledge that calcification should always be in close proximity to anatomical surfaces, the main goal for Aim 2 is to encourage effective sharing of imaging features from Aim 1 to also locate calcification. A novel loss for anatomical consistency will also be developed as part of this aim. Upon successful completion of this proposal, the final unified deep learning model will be able to use pre-operative 3D CT images to generate fully functional patient-specific volumetric FE meshes for accurate and versatile TAVR simulations, at a rate of ~20ms per image. This is a speed-up of several orders of magnitude compared to the current workflow, and thus will significantly accelerate biomechanics studies and bring FE simulations closer to clinical use. This work will be conducted at Yale University’s Biomedical Engineering department with guidance from Dr. James Duncan and Dr. Wei Sun under the F31 fellowship. The training will include extensive research at the intersection of biomedical image analysis, biomechanics, and machine learning, with emphasis on impactful clinical applications.