Fig. 3

a Schematic of the content-aware restoration deep-learning pipeline. Pairs of images were simultaneously collected with a beamsplitter in an epifluorescent microscope at different exposure times to create a low and high-SNR training dataset. b After subpixel registration, a deep neural network is trained to restore the test image from the high-SNR ground truth. c Structural similarity (SSIM) and Normalized Root Mean Squared Error (NRMSE) of the predicted images vs the ground truth (see also Supplementary Fig. 1). d The trained network is then used to restore low-SNR bioluminescent images. Scale bar = 50 μm. e, f CARE recovers high frequency components from degraded images. Estimation of the smallest resolvable feature from the noisy fluorescence images (e), Ground truth and deep-learning predictions and the f bioluminescent images. Fluorescence images were acquired with a ×10/0.3 objective and bioluminescence images were acquired with a ×40/1.25 silicon oil objective. (i) Average azimuth intensity of the Fourier-transformed image, plotted against frequency in the 2D power spectrum image. Mean ± SD (ii) Violin plot of the smallest resolvable feature (N = 9 and 15 technical replicates (images) for each condition, respectively). Black line is the median of the distribution. Inset show the synthetic PSF calculated assuming 470 nm emission light using the Born& Wolf method⁸¹. Theoretical resolution limit was calculated according to 0.61λ/NA_objective. Scale bar = 2 μm.