Refined Deep Layer Aggregation for Multi-Disease, Multi-View & Multi-Center Cardiac MR Segmentation

Di Liu, Zhennan Yan, Qi Chang, Leon Axel, Dimitris N. Metaxas

Abstract: Efficient and accurate segmentation of the heart is important for analysis of cardiac magnetic resonance imaging (MRI). Although many convolutional neural networks (CNNs) have been proposed to address cardiac segmentation in cine MRI, the task is still an open and challenging problem due to highly complex and variable cardiac shape in various pathologies and the ill-defined borders, particularly near the base and apex of the heart. Existing methods typically only segment the heart on either short-axis images or long-axis images without employing any complementary information from multi-view images for better segmentation of the ventricles. This can be problematic in the basal region, where the ventricles can be easily confused with the atria. In this paper, we propose a novel framework to jointly segment the ventricles in both short- and long-axis images. The method has two stages: 1) segment the two views independently and then 2) refine their segmentations by fusing the complementary information from the other views. The proposed method was evaluated on the MICCAI 2021 Multi-Disease, Multi-View & Multi-Center Cardiac MR Segmentation Challenge (M&Ms-2) dataset. The result shows improvement on the segmentation of the left and right ventricular cavities and the myocardium for both short-axis and long-axis cardiac MR images compared to the conventional state-of-the-art segmentation methods.