This study aimed to explore the use of Foundation Models (FM) for skeletal muscle segmentation from magnetic resonance (MR) images, which is a critical step in extracting morphological and functional biomarkers in neuromuscular disorders (NMDs). The research addresses challenges such as fat infiltration and ambiguous muscle boundaries by comparing the performance of traditional convolutional neural networks (CNNs) and emerging FMs like the Segment Anything Model (SAM) and its medical adaptation, MedSAM. A dataset of MR thigh images from 76 NMD patients, splitted in Early, Moderate, and Severe according to the degree of fat infiltration, was annotated for 12 muscle groups across 152 volumes. Fine-tuned SAM and MedSAM configurations (encoder-decoder and decoderonly) were evaluated alongside 2D and 3D nnU-Net CNN models. Performance was assessed using Dice Similarity Coefficient (DSC), Average Symmetric Surface Distance (ASSD), and 95% Hausdorff Distance (HD95). Additionally, uncertainty quantification (UQ) metrics, including Expected Calibration Error (ECE) and Negative Log-Likelihood (NLL), were employed to assess model reliability. SAM's finetuned encoder-decoder configuration achieved segmentation accuracy comparable to state-of-the-art 3D nnU-Net models (DSC: 0.925, 0.883, and 0.857 for Early, Moderate, and Severe cases) and significantly outperformed 2D nnU-Net model. SAM achieved superior calibration, particularly in the Severe group (ECE: 0.032 vs. 0.071 of nnU-Net 3D). Deep ensemble methods further improved segmentation reliability and UQ assessment. MedSAM did not surpass the SAM performance. In conclusion, SAM has demonstrated accurate and better calibrated segmentation, demonstrating its potential for medical imaging applications in challenging NMD datasets.