Generative Adversarial Semi-Supervised Network For Medical Image Segmentation

Abstract

Due to the limitation of ethics and the number of professional annotators, pixel-wise annotations for medical images are hard to obtain. Thus, how to exploit limited annotations and maintain the performance is an important yet challenging problem. In this paper, we propose Generative Adversarial Semi-supervised Network(GASNet) to tackle this problem in a self-learning manner. Only limited labels are available during the training procedure and the unlabeled images are exploited as auxiliary information to boost segmentation performance. We modulate segmentation network as a generator to produce pseudo labels whose reliability will be judged by an uncertainty discriminator. Feature mapping loss will obtain statistic distribution consistency between the generated labels and the real ones to further ensure the credibility. We obtain 0.8348 to 0.9131 dice coefficient with 1/32 to 1/2 proportion of annotations respectively on right ventricle dataset. Improvements are up to 28.6 points higher than the corresponding fully supervised baseline.