Title: Automated Reconstruction of Personalized 3D Myocardial Infarct Shapes from Cine MRI for Cardiac Digital Twins

Abstract:

Establishing precise 3D geometric models of myocardial infarction (MI) is a critical prerequisite for developing cardiac digital twins (CDTs) capable of accurately simulating electrophysiological dynamics associated with infarcts. While late gadolinium enhancement magnetic resonance imaging (LGE MRI) serves as the clinical gold standard for identifying MI locations, its dependence on contrast agents poses limitations for patients with renal impairment and hinders frequent longitudinal monitoring. In contrast, contrast-free cine MRI offers a viable alternative by visualizing abnormal ventricular wall motion, a strong indicator of infarcted tissue. This paper introduces a novel explicit geometry-motion embedded model designed to fully automatically reconstruct personalized, simulation-ready 3D MI geometries directly from multi-view cine MRIs. Our approach involves constructing a 4D (3D + time) biventricular mesh to explicitly isolate and decouple geometry-aware features from motion-aware features. Additionally, we implement a dual-branch module that enables adaptive fusion of geometry and motion data, thereby capturing the spatiotemporal dependencies necessary for mapping the infarcted region. To ensure biophysically consistent reconstructions, we employ multi-scale supervision guided by an AHA-17 segment-based cross-attention mechanism. Evaluation on a dataset of 225 cine MRIs revealed that the proposed method for 3D MI reconstruction achieves high accuracy, yielding an average Dice score of 0.678 $\pm$ 0.011. Downstream in-silico electrophysiological simulations confirmed that the results align closely with LGE-derived ground truths, underscoring the significant potential of this model for contrast-free scar characterization and its seamless incorporation into CDT frameworks. The source code will be made publicly available upon the manuscript’s acceptance for publication.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC