Dual-stream Generative Network Based Staining Transfer for Biomarker in Breast Cancer

Pathological examination is a crucial standard in cancer diagnosis, with breast cancer being one of the leading causes of morbidity and mortality in recent years, posing a major threat to health. Enhancing pathological examination capabilities has become an important way to save lives and improve patients’ quality of life. Common pathological examination methods include Hematoxylin and Eosin (H&E) and Immunohistochemistry (IHC) staining. H&E-stained images alone are often insufficient for cancer diagnosis, while IHC provides more comprehensive information for confirmed diagnosis. To address the challenges of limited IHC resources and high-cost consumption, we aim to generate virtual IHC images from H&E-stained images. In practice, it is difficult to perform multiple stains on the same tissue section, making it hard to obtain pixel-level matched data. To overcome this, we propose a dual-stream generative network that leverages pathological consistency constraints and a pathological representation network to extract pathological information and improve prediction accuracy. The network also incorporates structural similarity constraints and skip connections to enhance structural similarity. Additionally, we use stain unmixing results as annotated data, significantly reducing the workload of pathologists. We also conducted experiments to compare our model with models of similar functionality. In terms of pathological correlation, we have a lower Integrated Optical Density (IOD) and a higher Pearson-R, which are approximately 7.3% lower and 12.7% higher, respectively than the model with the best test results, so we have a higher pathological correlation. In terms of image quality, our Structural Similarity Index (SSIM) is higher than existing models, improving by approximately 8.8% compared to the model with the best test results, with higher image quality. These experiments show that our method has better stability and performance than the existing methods.