ScarNet Hybrid Transformer-U-Net for Myocardial Scar Segmentation in LGE Cardiac MRI
ScarNet is a hybrid deep-learning model for automated myocardial scar quantification from late gadolinium enhancement (LGE) cardiac MRI. Its novel contribution is coupling a MedSAM (Segment Anything-derived) transformer encoder with a U-Net convolutional decoder and tailored attention blocks so global context and fine boundary detail are fused for precise scar delineation in the left-ventricular myocardium. On 184 held-out ischemic-cardiomyopathy test patients it reached a median Dice of 0.912, far above fine-tuned MedSAM (0.046) and nnU-Net (0.638), with Monte Carlo noise-perturbation experiments confirming robustness and low bias versus manual expert segmentation.
ScarNet: A Novel Foundation Model for Automated Myocardial Scar Quantification from LGE in Cardiac
ScarNet is a hybrid deep-learning foundation model for automated quantification of myocardial scar from late gadolinium enhancement (LGE) cardiac MRI. It couples a MedSAM (Segment Anything-derived) t…