Fig. 2: Workflow and validation of 3Snet-CLID using standard structures.

a Training strategy of 3Snet-CLID. The denoised clear images generated by 3S-denoising are utilized as ground truth for training supervision; Noisy images with different noise levels are obtained by averaging ON images with varying frames (1, 2, 4, 8, 16), and then two noisy images at each noise level are randomly selected as a pair to train the self-supervised network for every training iteration. b Trained 3Snet-CLID network. A single-frame noisy image is denoised, subsequently upsampled, and applied to accomplish SR imaging. c Two adjacent FluoSpheres beads with different sizes were effectively resolved using 3Snet-CLID. Top: Images of 100 nm, 40 nm, and 20 nm beads under WF, PURE+deconvolution, DnCNN+ deconvolution, Sparse+deconvolution, and 3Snet-CLID; Bottom: Intensity profiles along the lines indicated by the arrow heads in Top row head. The peak distances between two adjacent beads with different sizes resolved by 3Snet-CLID are 126, 65, and 66 nm, while full width at half maximum (FWHMs) of corresponding WF intensity profiles are 310, 254, and 251 nm, respectively. Scale bar: 200 nm. d 3Snet-CLID has achieved a spatial resolution of 60 nm on the commercial Argo-SIM under WF microscopy. The intensity profiles of the corresponding double-line pairs (30 nm, 60 nm, 90 nm, 120 nm, 150 nm, 180 nm, and 210 nm distances) are displayed. Representative data from three biologically independent replicates are shown. Scale bar: 500 nm. e Cross-validation of 3Snet-CLID and thunderSTORM. Left: DNA origami samples labeled with paired fluorophores of 60 nm distances imaged by WF, thunderSTORM, and 3Snet-CLID; Middle: Zoomed-in regions of yellow box; Right: Intensity profiles along the lines indicated by the arrow heads and the average distance measured using two-peak Gaussian fitting function. Data are presented as mean values ± SD. n = 7 from three biological repeats. Scale bar: 500 nm.