Unsupervised domain adaptation network for medical image segmentation with generative adversarial networks

Deep learning algorithms have shown remarkable efficacy in segmenting medical images, but their performance deteriorates when encountering new test data due to domain shift. Given certain factors, unsupervised domain adaptation holds great significance in research on medical image segmentation. In this paper, we propose an unsupervised domain adaptation method for medical image segmentation by designing an improved generative adversarial network. We introduce a discriminator to supervise the output of the generator with a new class-sample adversarial loss, which enhances the credibility of the target domain samples generated in our network. Comprehensive experiments were performed on two fundus image datasets that are publicly accessible. Compared to the classic CycleGAN model, our strategy obtained an average increase of 2% in Dice and an average decrease of 3% in ASD. The experimental data supports the effectiveness of our approach.