Fig. 2

Fig. 2 Infrared-mediated image restoration improves image quality of degraded GFP images.

a, Single GFP and IR images (left) extracted at increasing detection depth in a 96 hpf Tg(h2b:GFP) zebrafish larva restored with either IR² or Noise2Void (N2V) (right). Scale bar, 100 µm. b, Example patches for the same zebrafish dataset shown in a arranged by increasing cell density. Scale bar, 5 µm. c, Violin plots of information content gain (relative to GFP) in patches extracted from the ground-truth (IR, dark red), IR²-reconstructed (IR², orange) and N2V-reconstructed (N2V, yellow) images. Vertical gray lines indicate s.d. d, Pearson correlation and SSIM obtained for patches extracted in the GFP, IR²- and N2V-restored images when compared to the ground-truth image (IR). e, SSIM relative to IR image as a function of detection depth for patches extracted throughout the sample. Data are presented as mean ± s.d. f, Single z planes at increasing detection depth for a Tg(His2AV-GFP) fly larva (8 hpf) extracted from the input (GFP) and ground-truth image (IR), as well as from restored images obtained from IR² and Noise2Void. Scale bar, 100 µm. g, Example patches for the same fly dataset shown in f. White asterisks indicate patches where artifacts were introduced or features were not reconstructed by the Noise2Void network. Scale bar, 5 µm. h, Violin plot of information content gain (relative to GFP) in patches extracted from the ground-truth (IR, dark red), IR²-reconstructed (IRIR, orange) and N2V-reconstructed (N2V, yellow) images. Vertical gray lines indicate s.d. i, Pearson correlation and SSIM relative to IR image, for GFP, infrared-mediated and Noise2Void reconstructions. j, SSIM relative to NIR images as a function of detection depth. Data are presented as mean ± s.d. In all violin plots, dashed lines represent 25/50/75th quartiles.