Exploring foundation models for multi-class muscle segmentation in MR images of neuromuscular disorders: A...

Deep learning (DL) models have shown promise for skeletal muscle (SM) segmentation in MR images, which is crucial for extracting biomarkers in neuromuscular disorders (NMDs). However, to ensure safe clinical use, models should provide uncertainty estimates, allowing radiologists to assess predictions and intervene when needed. Foundation Models (FMs) have the potential to play a significant role due to their strong generalization capabilities and well-calibrated predictions. However, their applicability in this context has not yet been explored. This study aims to develop an accurate and trustworthy technique by fine-tuning FMs to delineate fatty-infiltrated SM fascicles in NMD patients.

We fine-tuned Segment Anything Model (SAM) and MedSAM using two configurations - encoder/decoder and decoder only - and compared their performance against state-of-the-art 2D and 3D nnU-Net using a dataset of thigh MR images from 76 NMD patients, categorized into Early, Moderate, and Severe fatty infiltration groups. Accuracy was evaluated using the Dice Similarity Coefficient (DSC), while Uncertainty Quantification (UQ) was evaluated using the Expected Calibration Error (ECE) and the Negative Log-Likelihood (NLL). Deep Ensembles were used to convey epistemic uncertainty in addition to the aleatoric counterpart.

SAM's fine-tuned encoder/decoder outperformed nnU-Net 3D in Moderate and Severe cases (DSC: 0.886 vs 0.883 and 0.857 vs 0.850) and was comparable in Early (DSC: 0.925). MedSAM did not show an advantage over SAM. Regarding UQ, SAM exhibited superior calibration in Moderate and Severe groups (ECE: 3.6% vs. 5.1% and 3.3% vs. 7.1%), CONCLUSIONS:: In conclusion, our findings demonstrate that fine-tuning SAM yields superior performance, considering both accuracy and UQ metrics, highlighting its enhanced reliability in challenging NMD imaging scenarios.