Late Gadolinium Enhancement (LGE) cardiac MRI (CMRI) is the gold standard for assessing myocardial viability, but myocardial scar quantification is challenging due to inter-observer variability. While deep learning methods can achieve high agreement within a dataset, their generalization and clinical utility remain underexplored. In this work, we develop an LGE scar quantification model trained on pseudo-ground truth masks from expert-generated segment-wise labels. The dataset includes 159 and 53 patients for training and testing, respectively. To enhance performance, the model is initialized with a CMRI foundation model pretrained on 36 million images. The model obtained an AUC of 0.96 and accuracy of 0.89 for LGE detection. The trained model is then applied to an independent dataset of 662 patients from a different institution, without any modification. Here, we demonstrate the utility of the calculated scar burden as a prognosticator for detecting myopathologies (normal, dilated, hypertrophic, ischemic), over and above other cine-based measurements, by improving AUCs by 1–6% for various disease classes. This study demonstrates that a scar segmentation model trained without expert-annotated pixel-wise ground truth masks can generate clinically meaningful scar burden measurements, offering valuable diagnostic insights.