Normalization of HE-stained histological images using cycle consistent generative adversarial networks

摘要

In both histology and surgical pathology, the inherent individual appearance of the considered object on the one hand or the different staining protocols on the other hand must be compensated in addition to factors that influence the image acquisition (e.g scanning devices). This demand applies to hematoxylin-eosin (HE) staining being the standard method in pathology but also to all other histochemical and immunohistochemical staining. Regarding HE-staining, solutions and protocols are standardized at first glance. However, even within a single institution, protocols may vary slightly and may not be coordinated with other institutes. Especially when training deep neural networks, for example for image classification, there is a need for stain normalization of images so that models are transferable to other data sets.