Understanding Virtual Staining with generative adversarial networks for Osteoclast Imaging

Abstract

Virtual staining with generative adversarial networks is an efficient, non-invasive and scalable alternative to conventional cell staining, minimizing the need for destructive and time-consuming protocols. In this study, we investigate the explainability of a network trained to virtually stain osteoclast cultures, using intensity-based label-free input images. The model enables analysis of cell cultures without immunostaining. Explainability assessments, including receptive field and feature map analyses, show that the background in input images significantly influences staining predictions within cellular regions and the trained network performs an internal segmentation during the image transformation process. This suggests that contextual cues beyond cell boundaries are implicitly learned and integrated during training. By eliminating repetitive staining procedures, virtual staining enables longitudinal studies, allows multiplexing of individual samples, and reduces reagents and laboratory waste. Our findings enhance understanding of the virtual staining process and highlight its potential for biomedical research applications.