SCMR-ISMRM Workshop: 05 - Scientific Abstracts (AI and Machine Learning & Workflow)-Lecture 2 - Using machine-learning for fully automatic LGE scar quantification in the large multi-national Derivate Registry

Lecture 2 - Using machine-learning for fully automatic LGE scar quantification in the large multi-national Derivate Registry

Back to course

Video Transcription

Video Summary

The presentation describes a machine learning method for fully automatic quantification of late gadolinium enhancement scar in multicenter cardiac MRI. Using 573 post-infarct cases from a large international registry, the team trained a fully convolutional U-Net with data augmentation to segment healthy myocardium, dense scar, and non-dense scar. Performance was assessed with Dice, correlation, and Bland-Altman analysis. Results showed strong agreement with human ground truth, with only small bias and acceptable confidence intervals. The method works across 20 centers and may reduce observer variability, saving time and improving clinical scar assessment.

Keywords

late gadolinium enhancement

cardiac MRI

U-Net segmentation

myocardial scar quantification

machine learning