Abstract 4367687: Development of Personalized Myocardial Surface Mesh Models with LGE Scar Integration: a Pipeline for Machine Learning and Digital Twins

Abstract

Introduction: Advances in cardiac imaging offer new insights into the electrophysiologic correlates of myocardial fibrosis and scarring. Applications taking advantage of cardiac magnetic resonance (CMR) imaging are gaining traction, such as ventricular arrhythmia ablation guided by a substrate-based approach. However, a robust methodology to directly apply substrate level information from CMR imaging in conjunction with the planning and execution of electrophysiology studies is needed. This work addresses the development of 3D surface mesh models representing myocardial surfaces that integrate late gadolinium enhancement (LGE) scar from CMR to facilitate downstream risk stratification, pre-procedural planning, and intra-procedural navigation. Research Questions: Can we develop a 3D surface mesh model of myocardial surfaces that integrates LGE scar using CMR imaging? Methods: We generated voxel masks for myocardium, scar, blood pool, and empty space from CMR images utilizing an approach combining deep learning segmentation models and manual verification. Once segmentation masks were created, we delineated endocardial and epicardial surfaces by algorithmically identifying myocardial voxels adjacent to blood pool and empty space, respectively, across multiple short axis planes. For each plane, point-clouds representing these surfaces were extracted by ray casting. Robust polynomial functions were fitted to these point-clouds, and then used to create smooth curve segments. Endocardial and epicardial surfaces were created using curve segments as splines, which are subsequently refined and smoothed. Finally, scar LGE data, initially as a voxel mask, was converted to a raw 3D mesh using the marching cubes algorithm. Its vertices were then clustered and projected onto the myocardial wall to offset shifts that occurred from polynomial fitting. Results: This computational pipeline resulted in the successful creation of 49 personalized myocardial surface meshes with integrated LGE scar. These models provide a robust method for generating detailed 3D representations from patient-specific CMR images. Conclusion: Using a computational approach, we developed a robust method to generate 3D myocardial surface models with integrated LGE scar. These personalized myocardial surface models serve as an invaluable tool for enhancing risk stratification, pre-procedural planning, and potentially improving intra-procedural navigation for ventricular arrhythmia ablation.