Extended Data Fig. 3: Comparison of coherent LLS-SIM-based VSI-SR scheme with photo-reassignment-based laterally isotropic SR reconstruction.

a, In the training phase, raw SIM data were first collected with 6-phase illumination and 3-phase lattice illumination by each approach, and then reconstructed by applying the digital photo-reassignment algorithm and SR-SIM reconstruction algorithm, respectively, to generate 1D SR training data. The detection NA was set to 1.0 for both approaches, while the effective excitation NAs of structure illumination generation were set to 0.4 for LLS-SIM and 1.0 for incoherent photo-reassignment. Although using lower excitation NA, LLS-SIM achieved larger spatial frequency extension and finer structural details compared with those of photo-reassignment. b, In the inference phase, different neural network models, that is, RCAN and VSI-SR, were trained, which were both used to predict 1D SR outputs along multiple orientations from the diffraction-limited inputs. The 1D SR outputs were further assembled to produce the lateral isotropic reconstruction using Fourier projection or a generalized Wiener filter. The MIP of spatial profiles and corresponding Fourier spectrum are presented for each step, and a magnified inset and the line profile are displayed to demonstrate the resolution comparison of the two methods. Scale bar, 1 μm (excitation patterns), 5 μm (full FOV images) and 2 μm (magnified regions).