Generative Adversarial Networks for three-dimensional multi-channel virtual fluorescent staining

Fluorescent staining plays a crucial role in clinical pathology and life-science research for visualizing tissue and cellular structures. However, traditional staining methods are hindered by cost, time, and environmental impact. Deep learning has emerged as a transformative approach to address these challenges by computationally generating virtually stained images from label-free samples. We introduce a Wasserstein Generative Adversarial Network (WGAN) for three-dimensional fluorescent staining, leveraging multi-channel data to exploit the positive and negative colocalization information of cellular structures. The proposed model surpasses conventional U-Net architectures, producing sharper and more realistic images. Additionally, we designed a classification perceptual metric to evaluate the realisticness of the generated images that better correlates with expert evaluation.