Weakly supervised learning for scar reconstruction in personalized cardiac models: Integrating 2D MRI to 3D...

Accurate myocardial scar reconstruction is crucial for understanding arrhythmical risk and guiding treatment strategies in cardiac electrophysiology. Beyond providing anatomical insight, these reconstructions also serve as a foundational input for constructing patient-specific 3D computational models, which are increasingly used to simulate electrical activity and predict therapeutic outcomes. However, traditional interpolation methods struggle with the low inter-slice resolution (typically 8-10 mm) of late gadolinium-enhanced cardiac MRI (LGE-CMR) data, despite the high in-plane resolution (1-2 mm), leading to inaccuracies in the representation of scar morphology. In this study, we propose a deep learning-based interpolation framework that utilizes coordinate-based learning and Monte Carlo Dropout uncertainty estimation to reconstruct myocardial scar regions from sparse LGE-CMR slices. While the underlying methods are established, their application to myocardial scar reconstruction for use in personalized cardiac modeling represents a novel contribution. We evaluate our method on two different left ventricular (LV) models: a pre-clinical pig LV high-resolution (1.2 mm) gold-standard MRI, and a clinical LV with a more complex scar morphology. Our DL framework is trained using 3D anatomical coordinates as input, with a neural network architecture consisting of two hidden layers and dropout layers for uncertainty modeling. To improve consistency between 2D MRI data and 3D LV models, we introduce a custom loss function that enforces spatial and anatomical constraints. Experimental results demonstrate that the DL-based approach significantly outperforms the Log-Odds method in terms of scar segmentation accuracy, volumetric difference, and Dice similarity coefficient (DSC). The DL model trained with uncertainty estimation achieves the best performance, with MSE reduced to 0.094, DSC improved to 0.83, and volumetric error minimized to -2.72%. Additionally, uncertainty-aware training enhances the prediction of border zones between healthy and scarred myocardium, where traditional methods exhibit high errors. These findings highlight the effectiveness of deep learning-based interpolation for low-resolution LGE-CMR scar reconstruction, demonstrating its potential for improving patient-specific computational cardiac modeling. By incorporating uncertainty estimation and anatomical constraints, our approach provides a more accurate and clinically meaningful representation of myocardial scar morphology, paving the way for enhanced risk stratification and treatment planning in cardiac electrophysiology.