Fig. 4. Virtual fluorescence labeling without paired training data.

a Phase-contrast image of differentiated human motor neurons. The original image stack was average-projected for display. b Fluorescence images predicted by UTOM. c Ground-truth image acquired with a spinning disc confocal microscope. Axons, dendrites, and nuclei labeled with different fluorescence indicators are imaged in the red, green and blue channels, respectively. d Fluorescence images predicted by a multi-scale CNN trained on pixel-registered image pairs¹¹. The scale bar represents 200 μm. e–g Enlarged views of detailed structures extracted from the red, green, and blue channels, respectively. Axons, dendrites, and nuclei can all be resolved by UTOM