Conventional histopathology uses chemical staining as the gold standard for tissue analysis, but it is a time-intensive, laborious, and irreversible process. This study systematically evaluates the potential for deep neural networks in the virtual staining of tissue images obtained with regular brightfield microscopy. For tissues from multiple organs, we performed quantitative and visual evaluation of the reproduction accuracy of virtual staining vs. H&E-stained ground truth. Using variants of the generative adversarial network model pix2pix, we show that increasing neural network complexity can lead to higher virtual staining quality. Our study suggests that virtual staining could be used to reduce the need for chemical staining in histopathology.
【저자키워드】 deep learning, digital pathology, generative adversarial networks, H&E staining, generative AI, label-free microscopy, histological evaluation, unstained histology, unstained tissue imaging, virtual staining,