UGAN: Semi-supervised Medical Image Segmentation Using Generative Adversarial Network

Semi-supervised medical image segmentation is well known to solve the cost problem of medical data labelling. However, most methods are proposed for solving specific tasks, which means that a well-designed model is difficult to migrate to other datasets. It is a challenge to design a semi-supervised model adaptive to different datasets. We propose UGAN, i.e., generative adversarial network based on U-Net. Especially, the segmentation network can adjust itself to different tasks based on the signature of the dataset and obtain good segmentation results. We designed the discriminator to distinguish the ground truth from the segmentation results of the U - Net segmentation network. The results on the dataset ASOCA show the effectiveness of our network.