Unsupervised Multiple Virtual Histological Staining from Label-Free Autofluorescence Images
Lulin Shi, Ivy H. M. Wong, Claudia T. K. Lo, Lauren W. K. Tsui, Terence T. W. Wong · 2023
Clinical histopathological analyses usually require hematoxylin-and eosin-(H&E) as regular staining to visualize various tissue types and morphological changes, whereas some special stains are also essential to provide auxiliary information on particular components. However, it is infeasible to simultaneously implement diverse histological staining on the same tissue section. In this paper, we propose a multiple histological staining model that enables arbitrary staining image generation from label-free autofluorescence images. We use AdaIN to fuse styles into the image reconstruction process for source image content preservation. Moreover, direct image match loss is proposed to replace image reconstruction loss. Experimental results on mouse kidney tissue demonstrate the efficiency and advantage of our model compared to the baseline frameworks. Furthermore, we also validated the superior performance of the proposed model using mouse liver and heart tissues, which confirms that our method is generally applicable to multiple organs.