Motion and Blood Flow Analysis of Cardiovascular Magnetic Resonance Imaging using Deep Learning

Abstract

Background: Cardiovascular magnetic resonance (CMR) can provide precise quantitative information on tissue and blood motion, which enables assessment of mechanical function, tissue properties, and hemodynamics. Two such imaging techniques are CMR tagging and four-dimensional flow magnetic resonance imaging (4D Flow MRI). CMR tagging is a signal magnitude-based technique that enables measurement of cardiac function from regional deformation, while 4D Flow MRI is a phase-contrast imaging technique capable of measuring three-dimensional velocity information through time. However, these techniques have limitations such as low resolution and require extensive post-processing. A robust, efficient, and automated approach for motion quantification and post-processing would offer significant improvement for patient evaluation. Advancements in deep learning in recent years have shown promise in the field of computer vision, and can be adopted to medical image analysis. Aim: This thesis aimed to develop deep learning methods applicable to CMR images to provide quantitative improvements related to tissue motion and hemodynamics. Specifically, (1) strain estimation based on myocardial motion from MR tagging images, (2) improving spatial resolution of 4D Flow MRI in the aorta and cerebrovasculature, and (3) improving accuracy of wall shear estimation from 4D Flow MRI. Methods and Results: We developed several deep learning models, including (1) a network to estimate myocardial strain in CMR tagging, derived from multiple landmarks’ displacement through time. The method resulted in unbiased estimates with reasonable precision, suitable for robust evaluation in a high-throughput setting, (2) a super-resolution network to improve spatial resolution of 4D Flow MRI, trained on synthetic MRI data, generated from computational fluid dynamics simulations, and (3) a network to estimate aortic wall shear stress from 4D Flow MRI. Conclusion: Deep learning methods provided promising accuracy and performance for specific tasks, given enough training data. While the use of actual data is ideal for training, suitable ground truth can be difficult to obtain. Given sufficient similarity with actual data, synthetic data can be used as an alternative to train a deep learning model. More importantly, deep learning methods enable robust and more accurate measurements for patient evaluation.