Comparison of a Hybrid Mixture Model and a CNN for the Segmentation of Myocardial Pathologies in Delayed Enhancement MRI.

DE-MRI provides a reliable and accurate imaging technique for the assessment of pathological alterations in myocardial tissue. The clinically applied thresholding techniques enable the assessment of the amount of diseased tissue. To also assess distribution patterns, transmurality and micro-vascular obstruction, more accurate segmentation methods are needed. We compare a hybrid CNN and mixture model approach with a two single-stage U-Net segmentation: one based on the EMIDEC challenge data set, one with additional training data, and could achieve DICE coefficients of \(84.8 \% \), \(84.08 \%\), and \(82.95 \%\), respectively. We hope to further improve the promising results through an extension of the training set.