In and out of the nucleus: CNN based segmentation of cell nuclei from images of a translocating sensor

Tarek M Zikry, Katarzyna M Kedziora, Michael Rene Kosorok, Jeremy Purvis

Abstract: This study demonstrates application of convolutional neural networks (CNNs) for the analysis of a unique image analysis problem in fluorescence microscopy. We employed the U-Net CNN architecture and trained a model to segment nuclear regions in images of a translocating biosensor---which alternates between the nucleus and cytoplasm---without the need for a constant nuclear marker. The model provided high-quality segmentation results that allowed us to accurately quantify the extent of cyclin-dependent kinase activity in a population of cells. We envision that the development of this kind of analysis tools will enable biologists to design live-cell fluorescence imaging experiments without the need for providing a constant marker for a subcellular region of interest. As a consequence, they will be free to increase the number of biosensors measured in single cells or reduce the phototoxicity of cellular imaging.