Semi-supervised PR Virtual Staining for Breast Histopathological Images

Abstract

Progesterone receptor (PR) plays a vital role in diagnosing and treating breast cancer, but PR staining is costly and time-consuming, seriously hindering its application in clinical practice. The recent rapid development of deep learning technology provides an opportunity to address this problem by virtual staining. However, supervised methods acquire pixel-level paired H &E and PR images, which almost cannot be implemented clinically. In addition, unsupervised methods lack effective constraint information, and the staining results are not reliable sometimes. In this paper, we propose a semi-supervised PR virtual staining method without any pathologist annotation. Firstly, we register the consecutive slides and obtain the patch-level labels of H &E images from the registered consecutive PR images. Furthermore, by designing a Pos/Neg classifier and corresponding constraints, the output images maintain the Pos/Neg consistency with the input images, enabling the output images to be more accurate. Experimental results show that our method can effectively generate PR images from H &E images and maintain structural and pathological consistency with the reference. Compared with existing methods, our approach achieves the best performance.