Stain Style Transfer for Digital Pathological images of mouse liver Based on VCGAN

Stain normalization in pathology images is crucial for eliminating staining variations between different laboratories and enhancing the robustness of computer-aided diagnosis systems. This paper proposes a stain normalization method based on generative adversarial networks, called VCGAN. It employs Vision Transformer AutoEncoder (VTA) as the generator and integrates PatchGAN discriminator to achieve bidirectional translation between H&E and Masson-stained images. By introducing cycle consistency loss and identity loss, VCGAN effectively preserves the structural and stylistic consistency of images. Experimental results on the public Mouse Liver Stain Transfer (MLST) Dataset show that VCGAN has significant improvements compared to existing methods in terms of color consistency and the quality of generated stained images.