CS-Net: A Stain Style Transfer Network for Histology Images with CS-Gate Attention

Accurate segmentation of epithelium in Hematoxylin-Eosin (HE)-stained oropharyngeal cancer (OPC) pathological images is the premise for quantitative diagnosis. However, the pathological features of OPC are highly heterogeneous in morphology, while the appearance may be similar between epithelium and other tissues (such as stroma), which increases the difficulty of the segmentation task. To solve the above problem, we propose a two-step framework for epithelium tissue segmentation, including a stain-transfer step and an image-processing step. In the first step, we propose a stain-style transfer network called CS-Net, together with our proposed attention module called CS-Gate, to transform HE-stained images into Immunohistochemistry (IHC)-stained images, with the aim of enhancing the contrast among epithelium and other tissues. In the second step, we perform a series of image processing methods on the synthesized IHC-stained images to obtain the final binary mask of the OPC epithelium. The experimental results show that CS-Net can synthesize more stable images with a higher degree of restoration than the GAN-like networks, and the accuracy of the final obtained mask is also better than the current mainstream segmentation networks, reaching 92.48%. An external validation experiment shows that CS-Net has a stronger generalization capability.