A 3D deep learning approach based on Shape Prior for automatic segmentation of myocardial diseases

Accurate three-dimensional (3D) cardiac segmentation from late gadolinium enhancement (LGE)-MRI plays a critical role in designing a structure of reference for diagnosing many cardiac pathologies such as ischemia, myocarditis and myocardial infarction. This segmentation is however still a non-trivial task, due to the motion artifacts during acquisition, and heterogeneous intensity distributions. In this study, we develop a fully 3D automated model based on deep neural networks (DNN) for LGE-MRI myocardial pathologies (scar and No-reflow tissues) segmentation in a new expert annotated dataset. Considering that damaged tissue constitutes a small area of the whole LGE-MRI, we concentrated on myocardium region. Generated segmentation's approve that this preprocessing step promote the learning proceeding. Our proposed network, includes a shape reconstruction neural network which can be pre-trained and then be included into a combined loss function as a regularization term. To fix the volume size imbalance issue, we present a Multi-class Jaccard (IOU) based loss function to re-weight the training for all structures. Extensive experiments over the Automatic Evaluation of Myocardial Infarction from Delayed-Enhancement Cardiac MRI (EMIDEC) 2020 datasets, using diverse evaluation metrics such Hausdorff Distance (HD), Dice Similarity Coefficient (DSC), Absolute Volume Difference (AVD) and Absolute Volume Difference Rate according to volume of myocardium (AVDR) proved the promising performance of our approach, in segmenting various regions of myocardium. Our proposed algorithm yields to these significant Average Dice Coefficients over all predicted substructures, respectively : `Myocardium': 0.9507, `Infarctus': 0.7656, `No-reflow': 0.8377.