Random forest based left ventricle segmentation in LGE-MRI

Abstract

The leading cause of death worldwide is ischaemic heart disease. Late gadolinium enhanced magnetic resonance imaging (LGE-MRI) is the clinical gold standard to visualize regions of myocardial scarring. However, the challenge arises in the segmentation of the myocardial border, as the transition of scar tissue and blood pool can be very smooth, because the contrast agent accumulates in the damaged tissue and leads to various enhancements. In this work, a random forest based boundary detection approach is combined with a scar exclusion criterion. The final endocardial and epicardial border is found with the help of dynamic programming, which finds the distance weighted minimum through the boundary cost array. The segmentation method is evaluated using a 5-fold cross validation on 100 clinical LGE-MRI data sets. The Dice coefficient resulted in an overlap of 0.83 for the endocardium as well as for the epicardium.