Unpaired Stain Transfer Using Pathology-Consistent Constrained Generative Adversarial Networks

Pathological examination is the gold standard for diagnosis of cancer. Common pathological examinations include hematoxylin-eosin (H&E) staining and immunohistochemistry (IHC). In some cases, it hard to make accurate diagnoses cancer by referring only H&E images. Whereas, IHC can further provide enough evidence process. Hence, generation virtual images from H&E-stained will be a good solution current accessibility issue, especially low-resource regions. However, existing approaches have limitations in microscopic structural preservation consistency pathology properties. addition, pixel-level paired data available. our work, we propose novel adversarial learning method effective Ki-67-stained image corresponding image. Our takes fully advantage similarity constraint skip connection improve details preservation; representation network are first proposed enforce generated source hold same properties different domains. We empirically demonstrate effectiveness approach on two unpaired histopathological datasets. Extensive experiments indicate superior performance that surpasses state-of-the-art significant margin. also achieves stable unbalanced datasets, which shows has strong robustness. believe potential clinical advance progress computer-aided multi-staining histology analysis.