Voloshin, N.; Putlyaev, E.; Chechekhina, E.; Usachev, V.; Karagyaur, M.; Bozov, K.; Grigorieva, O.; Tyurin-Kuzmin, P.; Kulebyakin, K.

Microscopy is an essential method in modern biology, and brightfield microscopy methods (phase contrast, differential interference contrast, etc.) are being widely used and actively developed since they do not require sample fixation and staining. But they produce low contrast images, where cells have similar intensity to the background. In this work we developed and tested a set of deep learning object detection YOLO models that detect cell nuclei in phase contrast images, which allows cell count and tracking without staining. We created a large dataset consisting of more than 100,000 640x640 pixels images with more than 3 million nuclei of 4 different cell cultures (CHO, HEK293, iPSCs, and MSCs). Using images from various microscopes and cameras, as well as full-scale augmentations, we developed a set of highly generalized models that can detect nuclei in images across different cell types and imaging conditions, including different microscopes and contrast methods. Combined with sliced inference methods, these algorithms can be applied to images of any size, allowing studies of large quantities of cells. Moreover, we developed a training-free calibration and testing algorithm based on confidence threshold optimization. It allows for fine-tuning of models for specific cell types and/or imaging options and evaluating the accuracy of the calibrated model. This provides a highly controllable and reliable method for studying cell proliferation rate, single cell tracking and other scenarios. Additionally, we developed a NuclePhaser plugin for Napari (https://github.com/nikvo1/napari-nuclephaser), which allows users to calibrate, test and apply our models in code-free manner. Given that the YOLO models are fast and can run at sufficient speeds even on CPUs, this makes our work highly accessible to a wide range of researchers.