Abstract
Strain quantification and strain-based prediction play crucial roles in evaluating cardiac function, providing valuable clinical data for heart disease assessment, diagnosis, and treatment planning. Motion detection is fundamental to strain quantification, with deep learning approaches showing significant promise. However, current unsupervised methods have limited accuracy, particularly in low-contrast regions, while supervised approaches often fail to fully utilize underlying motion features. Beyond motion detection, the development of automatic strain-based prediction models remains a critical yet under-explored area, presenting unique challenges that require innovative solutions to advance clinical applications.
This dissertation addresses three key technical challenges: First, improving motion detection accuracy, especially in regions with low image contrast or complex motion patterns. Second, overcoming the limited availability of training data for developing automatic strain-based prediction algorithms. Third, distinguishing between misleading motion patterns that originate from distinct underlying conditions but have similar appearances, which existing methods often fail to differentiate.
Our key contributions include: (i) A latent motion diffusion model for accurate motion tracking from cine Magnetic Resonance Imaging (MRI), guided by high-accurate ground truth data from Displacement Encoding with Stimulated Echoes (DENSE) MRI during training, while requiring only cine MRI for inference; (ii) A comprehensive multi-task learning framework that improves prediction accuracy with limited training data by introducing an auxiliary task that provides complementary information, enhancing overall performance without requiring additional data; and (iii) A multimodal integration approach that enhances the model's robustness to misleading patterns by incorporating complementary information about underlying cardiac conditions from an additional imaging modality, without requiring this extra data during clinical inference. These advancements collectively represent a significant step forward in deep learning-based medical strain quantification and strain-based prediction, offering enhanced accuracy and efficiency. Our research paves the way for more accurate strain-based analysis in clinical settings, with potential applications in various medical domains.
