A Virtual Staining Method for Immunohistochemical Images of Breast Cancer

Breast cancer is a common malignant cancer. Detection of human epidermal growth factor receptor 2 (HER2) status based on immunohistochemistry (IHC) is an effective method for breast cancer diagnosis. Early diagnosis and treatment in breast cancer can greatly reduce the mortality rate, but the immunohistochemistry-based staining method is time-consuming and expensive. For the standpoint of saving labor, resources and time costs, direct generation of virtual IHC images based on hematoxylin and eosin (HE) slices is an efficient and accurate method. We proposed an IHC image generation method based on generative adversarial network (GAN). A content encoder, a style encode and a discriminator were used in our method, the first encoder of which extracted unstained images in the generator, the second encoder which obtained a stained style to stain unstained images to generate IHC images, and the discriminator of which is designed to judge the authenticity and the classification of HER2 expression at the same time. Experiments showed that our method outperforms other methods in PSNR and SSIM metrics in BCI dataset.