Fig. 3: Performance of DNN-2 for resolution recovery for in vivo head-fixed 2 P images acquired at various scanning densities (from 512 spirals/frame to 512/M spirals/frame, M = 2–32).
From: Deep-learning two-photon fiberscopy for video-rate brain imaging in freely-behaving mice
Each scanning density requires a separately trained network. a Representative original image (with a scanning density of 512 spirals/frame) and b the corresponding DNN-1 output image which served as the reference for quantitative image quality comparison. c Top-row: in vivo images of a lower scanning density as the input for testing DNN-2. These images were obtained by digitally down-sampling the original image by a factor of M along the radial direction; bottom-row: DNN-2 output images. d Structural details of a selected region of interest with respect to different M factors. e Reconstruction accuracy in terms of the multi-scale structure similarity (MS-SSIM) index and normalized root-mean-square error (NRMSE) for the entire testing image set with respect to the down-sampling factor M. The data are presented as mean values (data points) ± standard deviations (error bars). The measurements were made over the whole image stack with a total number of images n = 800. Scale bars in (a, b): 20 µm.
