Abstract

Although three-dimensional (3D) fluorescence microscopy is an essential tool for life science research, the fundamentally-limited optical throughput, as reflected in the compromise between speed and resolution, so far prevents further movement towards faster, clearer, and higher-throughput applications. We herein report a dual-stage mutual-feedback deep-learning approach that allows gradual reversion of microscopy degradation from high-resolution targets to low-resolution images. Using a single blurred-and-pixelated 3D image as input, our trained network infers a 3D output with notably higher resolution and improved contrast. The performance is better than conventional 1-stage network approaches. It pushes the throughput limit of current 3D fluorescence microscopy in three ways: notably reducing the acquisition time for accurate mapping of large organs, breaking the diffraction limit for imaging subcellular events with faster lower-toxicity measurement, and improving temporal resolution for capturing instantaneous biological processes. Combining our network approach with light-sheet fluorescence microscopy, we demonstrate the imaging of vessels and neurons in the mouse brain at single-cell resolution and with a throughput of 6 minutes for a whole brain. We also image cell organelles beyond the diffraction limit at a 2-Hz volume rate, and map neuronal activities of freely-moving C. elegans at single-cell resolution and 30-Hz volume rate.