UNPAIRED H&E TO PR STAIN TRANSFER WITH SELF-SUPERVISED AUXILIARY SEGMENTATION

Yiyang Lin, Yifeng Wang, Yang Chen, Zijie Fang, Zexin Li, Yongbing Zhang

Abstract:

Histochemical staining is a critical step in the diagnosis of
cancer, where hematoxylin-eosin (H&E) stain is used most
commonly in clinical practice. However, the H&E images
often cannot be used for making accurate diagnoses. To
this end, pathologists must perform immunohistochemical
(IHC) stain, which is time-consuming and costly. In the field
of computer-aided diagnosis, existing models can virtually
generate IHC staining images, but they often require pixelaligned data and annotations from pathologists, which are
difficult to be obtained. To address this problem, we propose
a self-supervised PR (a typical type of IHC) virtual staining
model utilizing unpaired data without pathologists’ annotations for the first time. Based on the observation that PR
images are easy to be segmented, we introduce segmentation
as the proxy task to make the virtual staining more accurate. Experimental results show that our model can generate
PR images with the highest accuracy. Moreover, our model
achieves the desired results on an external dataset.