Staining Cells Virtually Offers Alterative Approach to Chemical Dyes
Staining Cells Virtually Offers Alterative Approach to Chemical Dyes lead image
A commonly used technique for quantitative analysis of cellular structures involves chemical staining with fluorescent dyes. There are, however, several drawbacks to this approach, including limitations on the types and combinations of dyes that can be used, the invasive nature of the technique, and possible toxicities associated with the chemical agents and the fluorescence itself.
Helgadottir et al. describe an alternative approach using brightfield imaging in combination with a neural network that produces a type of virtual staining of cell structures. The investigators tested their method on a stack of brightfield images of human adipocyte, or fat, cells. They showed their method can produce virtually-stained images that compare favorably to fluorescently-stained images.
The neural network used in this work is a conditional generative adversarial neural network, or cGAN, and consists of two networks: a generator that produces virtually-stained images and a discriminator that determines whether the result is a fluoroscently-stained image or one created by the generator.
“The generator progressively becomes better at generating virtually-stained images that can fool the discriminator,” author Giovanni Volpe said. “In turn, the discriminator becomes better at discriminating chemically-stained images from generated images.”
With this method, the investigators were able to visualize several cellular structures, including lipid droplets, cell nuclei and regions of cytoplasm. They could determine quantitative biological information from these results by employing a feature extraction pipeline to calculate the number of cell structures in each image and other data.
Other researchers are invited to use this method by downloading an open source Python software package prepared by the authors. The code can be modified for specific applications in other labs.
Source: “Extracting quantitative biological information from brightfield cell images using deep learning,” by Saga Helgadottir, Benjamin Midtvedt, Jesús Pineda, Alan Sabirsh, Caroline B. Adiels, Stefano Romeo, Daniel Midtvedt, and Giovanni Volpe. Biophysics Reviews (2021) The article can be accessed at https://doi.org/10.1063/5.0044782 .
