Myocardial Edema and Scar Segmentation Using a Coarse-to-Fine Framework with Weighted Ensemble

In this work, we implement a deep learning-based segmentation algorithm that can automatically segment left ventricular (LV) blood pool, right ventricular (RV) blood pool, LV normal myocardium, LV myocardial edema and LV myocardial scar from multi-sequence Cardiac Magnetic Resonance (CMR) images. Since the edema and scar region is very small, we adapt a coarse-to-fine segmentation strategy that contains two segmentation neural networks. Firstly, we use a coarse segmentation model to predict the cardiac structure area especially the myocardium part where the scar and edema regions distribute. Then we use the fine segmentation model to get a detailed prediction for edema and scar regions. Finally, we apply a weighted ensemble model to integrate the prediction from 2D and 2.5D networks. Our proposed framework achieves an average Dice score of 0.64 for LV myocardial scar and 0.41 for LV myocardial edema on 5-fold cross validation dataset from myocardial pathology segmentation combining multi-sequence CMR(MyoPS) challenge, while achieving an average Dice score of 0.67 and 0.73 in LV myocardial scar and the union of scar and edema on test set, respectively.