2025/10 - A Novel Foundation Model for Automated Myocardial Scar Quantification from Late Gadolinium-Enhancement Images [Journal Club]-ScarNet: A Novel Foundation Model for Automated Myocardial Scar Quantification from Late Gadolinium-Enhancement Images

ScarNet: A Novel Foundation Model for Automated Myocardial Scar Quantification from Late Gadolinium-Enhancement Images

Video Transcription

Video Summary

The JCMR Journal Club opened with an introduction to a study on SCARnet, a novel foundation-model-based AI system for automated myocardial scar quantification from late gadolinium enhancement (LGE) cardiac MRI. The presenter explained that current scar segmentation methods are limited: CNNs/U-Nets capture local detail but miss broader context, while transformer-based foundation models such as MedSAM provide global context but lack cardiac-specific precision. SCARnet combines both approaches and adds a scar-attention block, then fine-tunes the model on a large multi-center dataset from the DETERMINE study (685 ischemic cardiomyopathy patients, 7,066 LGE images). Ground truth was created by expert cardiologists using manual full-width half-maximum contours.<br /><br />SCARnet outperformed MedSAM and U-Net variants, achieving closer agreement with manual segmentation, better robustness to noise, and fewer false positives/negatives, especially for small, irregular scars. Performance plateaued with relatively small training sizes because the model starts from strong pre-trained weights. The discussion covered clinical deployment challenges, including privacy, FDA regulation, and the need for broader validation. Future directions include testing on new datasets, handling poor image quality and microvascular obstruction, and integrating radiology reports with imaging for richer automated output.

Keywords

SCARnet

myocardial scar quantification

late gadolinium enhancement

cardiac MRI

foundation model

MedSAM

U-Net

DETERMINE study

scar segmentation