Late gadolinium enhancement (LGE) cardiac magnetic resonance (MR) imaging is a key technique for assessing myocardial damage in various cardiovascular diseases. Precise identification of cardiac structures is essential for computing clinical biomarkers used in diagnosis and prognosis. While many deep learning methods have been proposed for automatic segmentation, their generalizability is hindered by limited training data and domain adaptation issues, restricting clinical deployment. Foundation models have recently emerged as a promising solution, able to perform zero and few-shot segmentation across various imaging modalities through large-scale pretraining. Among them, the Segment Anything Model (SAM) has become a widely recognized reference, with adaptations for medical imaging, such as MedSAM and SAM-Med2D. In this study, we evaluate MedSAM and SamMed2D for segmenting the left ventricle and myocardium in LGE MR images from a private dataset comprising 135 patients. We first demonstrate that zero-shot performance remains limited, due to the scarcity of LGE MR data in their pretraining. Next, we show that fine-tuning the MedSAM decoder significantly improves segmentation quality, surpassing the nnU-Net baseline, though it requires precise bounding box initialization. We thus propose a modified MedSAM architecture that enables multi-class segmentation from a single bounding box. However, our experiments reveal that despite various improvements, MedSAM continues to produce mixed results. While our approach can segment multiple structures with one BB, it still requires accurate initialization, and its performance converges towards that achieved by nnU-Net.