Left Ventricle Segmentation in LGE-MRI: Filter Based vs. Learning Based

Ischaemic heart disease is the number one cause of death world wide, which is in close relation with heart failure. If patients suffer from drug-refractory heart failure with a reduced ejection fraction, cardiac resynchronization therapy is a treatment option. For planning the procedure, precise information about the left ventricle's anatomy and scar distribution is required. The clinical gold standard to visualize scar is late gadolinium enhanced magnetic resonance imaging (LGE-MRI). The challenge arises in the myocardium segmentation of these sequences which is a pre-requisite for an accurate scar quantification. In this work, we compare a filter based approach against a learning based approach for LGE-MRI segmentation. For both approaches the segmentation workflow consists of four major steps. First, the left ventricle is detected. Second, the blood pool is estimated. Third, the endocardium is refined using scar information. Fourth, the epicardium is extracted.The proposed methods were evaluated on 100 clinical LGE-MRI data sets. For the learning based approach a 5-fold nested cross-validation is applied to evaluate the hyper-parameters. The learning based segmentation achieves slightly better results, with a Dice score of 0.82 ± 0.09 for the endocard and 0.81 ± 0.08 for the epicard.