Jiarui (Jerry) Xing, PhD dissertation defense, University of Virginia.
Myocardial strain analysis and strain-based prediction play an important role in assessing, diagnosing, and planning treatment for heart diseases. However, existing methods often fall short in achieving automatic and accurate results, which limit their deployment in real-world clinical applications. This dissertation introduces novel deep motion networks for improved accuracy of strain quantification from cardiac magnetic resonance (CMR) image sequences, followed by a suite of strain-based prediction models to combat the critical challenges of manual processing, the scarcity of labeled training data, and the presence of misleading features in complex image domains. My key contributions entail in developing: (i) a supervised latent motion diffusion model that substantially improves the accuracy of motion prediction from time-series CMR images, enabling precise strain quantification; (ii) a comprehensive multi-task learning framework that improves strain-based prediction performance by incorporating auxiliary tasks to maximize the utilization of available training data; and (iii) a multimodal learning network that enhances model robustness against misleading motion features and strain patterns. These advancements represent a significant leap in CMR-based strain analysis, demonstrating efficiency and applicability in real-world clinical applications, specifically in cardiac resynchronization therapy, using routinely collected clinical image data.