ScarNet: a novel foundation model for automated myocardial scar quantification from late gadolinium-enhancement images

BACKGROUND: Late gadolinium enhancement (LGE) imaging remains the gold standard for assessing myocardial fibrosis and scarring, with left ventricular (LV) LGE presence and extent serving as a predictor of major adverse cardiac events (MACE). Despite its clinical significance, LGE-based LV scar quantification is not used routinely due to the labor-intensive manual segmentation and substantial inter-observer variability. METHODS: We developed ScarNet that synergistically combines a transformer-based encoder in medical segment anything model (MedSAM), which we fine-tuned with our dataset, and a convolution-based decoder in UNet with tailored attention blocks to automatically segment myocardial scar boundaries while maintaining anatomical context. This network was trained and fine-tuned on an existing database of 401 ischemic cardiomyopathy patients (4137 2D LGE images) with expert segmentation of myocardial and scar boundaries in LGE images, validated on 100 patients (1034 2D LGE images) during training, and tested on unseen set of 184 patients (1895 2D LGE images). Ablation studies were conducted to validate each architectural component's contribution. RESULTS: In 184 independent testing patients, ScarNet achieved accurate scar boundary segmentation (median DICE=0.912 [interquartile range (IQR): 0.863-0.944], concordance correlation coefficient [CCC]=0.963), significantly outperforming both MedSAM (median DICE=0.046 [IQR: 0.043-0.047], CCC=0.018) and nnU-Net (median DICE=0.638 [IQR: 0.604-0.661], CCC=0.734). For scar volume quantification, ScarNet demonstrated excellent agreement with manual analysis (CCC=0.995, percent bias=-0.63%, CoV=4.3%) compared to MedSAM (CCC=0.002, percent bias=-13.31%, CoV=130.3%) and nnU-Net (CCC=0.910, percent bias=-2.46%, CoV=20.3%). Similar trends were observed in the Monte Carlo simulations with noise perturbations. The overall accuracy was highest for ScarNet (sensitivity=95.3% (163/171); specificity=92.3% (12/13)), followed by nnU-Net (sensitivity=74.9% (128/171); specificity=69.2% (9/13)) and MedSAM (sensitivity=15.2% (26/171); specificity=92.3% (12/13)). CONCLUSION: ScarNet outperformed MedSAM and nnU-Net for predicting myocardial and scar boundaries in LGE images of patients with ischemic cardiomyopathy. The Monte Carlo simulations demonstrated that ScarNet is less sensitive to noise perturbations than other tested networks.