Whole Heart and Great Vessel Segmentation with Context-aware of Generative Adversarial Networks

Automatic segmentation of cardiac magnetic resonance imaging (CMRI) is an important application in clinical tasks. However, semantic segmentation of the myocardium and blood pool in CMR is a challenge due to differentiating branchy structures and slicing fuzzy boundaries. In this paper, we propose an automatic deep architecture for simultaneous myocardium and blood pool segmentation on patients with congenital heart disease (CHD). Inspired by vanilla generative adversarial networks (GANs), we propose a cascade of conditional GANs for semantic segmentation. The proposed cascade has three stages that are designed to share convolutional features and weights. Each stage has a conditional generative adversarial network with a unique loss function and trains on different images from the same patients. We further apply AutoContext Model to implement a context-aware generative adversarial network. The proposed method evaluated on the HVSMR dataset and the experimental results demonstrated the superior performance of our approach.