Recently, numerous adversarial learning based domain adaptation methods for semantic segmentation have been proposed. For example, Vu et al. minimized entropy of the prediction and also introduced the entropy discriminator to discriminate the source entropy maps from the target entropy maps. In 2018, Tsai et al. found output space contains rich information thus, they proposed the output space discriminator. Both of the methods have achieved promising results in street scene segmentation, while for medical image segmentation, we can take advantage of the information in the shape of the organs. For instance, point clouds can be used to create 3D models to incorporate shape representation as prior information. Cai et al. introduced the organ point network. It takes deep learning features as input and generates the shape representation as a set of points located on the organ surface. They optimized the segmentation task with the point network as an auxiliary task so that the shared parameters could benefit from both tasks. They also proposed a point cloud discriminator to guide the model to capture the shape information better.
We aim to combine the ideas from the previous works and investigate the impact of output space and entropy discriminators for multi-modality cardiac image segmentation. We want to employ point cloud classification as an auxiliary task, and introduce a point cloud discriminator to discriminate the source point cloud from the target point cloud.