Research and Application of Cell Image Segmentation Based on Generative Adversarial Network

Pixel-level cell segmentation plays an important role in cell image processing. However, very large amount of labeled images are always required for accurate segmentation. In this study, we propose a cell image segmentation algorithm based on generative adversarial network (CIS-GAN) using only small labeled images. A generator and a discriminator are included in our CIS-GAN. Specifically, the generator adopts a spatial pyramid pooling feature fusion module to predict the cell segmentation map. The discriminator applies a fully convolutional network based on ResNet to discriminate the probability distribution between the predicted segmentation map and the ground-truth label at each pixel. The adversarial learning between the generator and discriminator as well as the spatial cross entropy loss promotes the accurate segmentation of the generator with only small labeled images. The results on cell images from a certain type of spacecraft show that our CIS-GAN can obtain promising segmentation results with small amount of labeled images, and can achieve superior quantitative evaluation results (pixel accuracy, mean accuracy, and mean IU) and qualitative visual segmentation results than the state-of-the-art methods including FCN, Unet, SegNet and PSPNet.