Fig. 3: Characterizations and demonstrations of 3D ZS-DeconvNet.
From: Zero-shot learning enables instant denoising and super-resolution in optical fluorescence microscopy
a The network architecture of 3D ZS-DeconvNet and the schematic of its training phase. b The schematic of the inference phase of 3D ZS-DeconvNet. c Representative maximum intensity projection (MIP) SR images of F-actin, Mito outer membrane, and ER reconstructed by sparse deconvolution (second column), 3D ZS-DeconvNet (third column), and LLS-SIM (fourth column). Average sCMOS counts of the highest 1% pixels for raw images before processing are labeled on the top right corner. d Statistical comparisons of RL deconvolution, sparse deconvolution and ZS-DeconvNet in terms of PSNR and resolution on different specimens (n = 40 regions of interest). The resolution was measured by Fourier ring correlation analysis74 with F-actin image stacks. Center line, medians; limits, 75% and 25%; whiskers, maximum and minimum. Source data are provided as a Source Data file. e Time-lapse three-color 3D rendering images reconstructed via 3D ZS-DeconvNet of ER, H2B, and Mito, showing their transformations in morphology and distribution as well as interaction dynamics during mitosis (Supplementary Video 5). f Representative three-color images obtained with conventional LLSM (first column), sparse deconvolution (second column), DeepCAD based deconvolution (third column) (Methods), and 3D ZS-DeconvNet (fourth column). The comparisons are performed on two typical timepoints of the time-lapse data shown in (e). Scale bar, 5 μm (c, e, f), 1.5 μm (zoom-in regions of c), 2 μm (zoom-in regions of f).
