Towards Histopathological Stain Invariance by Unsupervised Domain Augmentation using Generative Adversarial Networks

The application of supervised deep learning methods in digital pathology is\nlimited due to their sensitivity to domain shift. Digital Pathology is an area\nprone to high variability due to many sources, including the common practice of\nevaluating several consecutive tissue sections stained with different staining\nprotocols. Obtaining labels for each stain is very expensive and time consuming\nas it requires a high level of domain knowledge. In this article, we propose an\nunsupervised augmentation approach based on adversarial image-to-image\ntranslation, which facilitates the training of stain invariant supervised\nconvolutional neural networks. By training the network on one commonly used\nstaining modality and applying it to images that include corresponding, but\ndifferently stained, tissue structures, the presented method demonstrates\nsignificant improvements over other approaches. These benefits are illustrated\nin the problem of glomeruli segmentation in seven different staining modalities\n(PAS, Jones H&E, CD68, Sirius Red, CD34, H&E and CD3) and analysis of the\nlearned representations demonstrate their stain invariance.\n