Research On Cell Image Segmentation Based On Generative Adversarial Network

The segmentation of cell images plays an important role in cancer detection.However,because cell images have complex background,such as overlapping,poor contrast,and disturbances of impurities,cell segmentation remains challenging even after years of research.For the traditional segmentation methods,they are difficult to segment multi-complex background images at the same time.Although the deep learning segmentation models can solve the problem of segmentation of multi-complex background images,they are hard to segment overlapping cells in the way of pixel classification.To solve these problems,a new deep learning segmentation model,called Cell-GAN,is proposed.As a generative adversarial network,Cell-GAN tries to realize the differentiation of image information by using the probability distribution of cell morphology learned by training,and then to achieve the segmentation with the help of autoencoder.For each cell to be segmented,CellGAN judges the integrity of a cell and treats other cellular information,except for overlapping parts,as the background.Finally,a single-cell image without background is generated and the contour of it is the final segmentation line.To distinguish each cell in the cell image,Cell-GAN takes dual-image input mode: the cell image and the relative single nuclear image which is named as guide factor.The guide factor is used to locate the cell to be segmented which is also the key to segment overlapping cells.Due to the consideration of computing resources,it is difficult for the deep learning segmentation model to segment large size images directly,hence,image cropping is necessary.In this paper,a novel cropping method,named recurrent image cropping(R-crop),is proposed,which crops images by making full use of the generation capability of the trained Cell-GAN.The cropping mechanism of Rcrop is to use the generated cell image of Cell-GAN to obtain accurate cell size,and then to crop the image and generate the cell image again.After that,the above process is repeated until the area fluctuation of the generated image tends to be stable,which means that the segmentation is completed.The combination of R-crop and Cell-GAN forms the first deep learning segmentation model RCell-GAN,which achieves to segment overlapping cells automatically.In this paper,two cell segmentation methods are used to compare the segmentation performance with RCell-GAN in self-built datasets and ISBI 2015 datasets.The segmentation results demonstrate that the proposed method has the best segmentation performance and excellent generalization capability.