Ziwen Liu, Eduardo Hirata-Miyasaki, Soorya Pradeep, Johanna V. Rahm, Christian Foley, Talon Chandler, Ivan Ivanov, Hunter Woosley, Tiger Lao, Akilandeswari Balasubramanian, Chad Liu, Manu Leonetti, Carolina Arias, Adrian Jacobo, Shalin B. Mehta
- Dynamic imaging of landmark organelles, such as nuclei, cell membrane, nuclear envelope, and lipid droplets enables image-based phenotyping of functional states of cells. Multispectral fluorescent imaging of landmark organelles requires labor-intensive labeling, limits throughput, and compromises cell health. Virtual staining of label-free images with deep neural networks is an emerging solution for this problem. Multiplexed imaging of cellular landmarks from scattered light and subsequent demultiplexing with virtual staining saves the light spectrum for imaging additional molecular reporters, photomanipulation, or other tasks. Published approaches for virtual staining of landmark organelles are fragile in the presence of nuisance variations in imaging, culture conditions, and cell types. This paper reports model training protocols for virtual staining of nuclei and membranes robust to cell types, cell states, and imaging parameters. We developed a flexible and scalable convolutional architecture, named UNeXt2, for supervised training and self-supervised pre-training. The strategies we report here enable robust virtual staining of nuclei and cell membranes in multiple cell types, including neuromasts of zebrafish, across a range of imaging conditions. We assess the models by comparing the intensity, segmentations, and application-specific measurements obtained from virtually stained and experimentally stained nuclei and membranes. The models rescue the missing label, non-uniform expression of labels, and photobleaching. We share three pre-trained models, named VSCyto3D, VSCyto2D, and VSNeuromast, as well as VisCy, a PyTorch-based pipeline for training, inference, and deployment that leverages the modern OME-Zarr format.
