Virtual immunohistochemistry by conditional generative adversarial networks

Histopathological diagnosis is crucial for assisting doctors and veterinarians with timely treatment strategy selection and optimal patient management. After the hematoxylin and eosin (H&E) staining process, immunohistochemistry (IHC) staining is often required, resulting in a longer turnaround time but providing more disease-specific information. In this study, we propose a Virtual IHC Generative Adversarial Network (VihcGAN) model to transform the H&E stained images of formalin-fixed sections from canine lymph nodes into CD3 and PAX5 IHC stained images, providing a feasible procedure to obtain virtually IHC stained images paired with H&E images. We further propose a novel metric stain Intersection over Union (IoU) to evaluate the accuracy of IHC stained images by incorporating the knowledge of IHC staining, such as the type, number, and position of stained cells. Virtual IHC staining via VihcGAN is computationally fast, with 1 second per image, and thus can bypass time-consuming and costly IHC procedures. Moreover, VihcGAN generates images with a high resolution of 2048 $\times$ 2048, which is much higher than related studies. Experimental results on image quality, metrics, expert grading, and visualization assessment demonstrate the effectiveness of our approach. Source code for VihcGAN and Stain IoU metric are available at https://github.com/coffeeNtv/VihcGAN .