Fig. 3: VsLFM shows robustness to noise and optical aberrations.

a, Orthogonal MIPs of 1-μm-diameter synthetic tubulins, acquired by sLFM with a ×63/1.4 NA oil-immersion objective in ideal imaging conditions, regarded as ground truth. b, VsLFM, VCD-Net and HyLFM-Net results after the input contaminated by strong mixed Poisson–Gaussian noise. We set the image bit depth to 16, the variance of Gaussian noise to 5, and the photon number of the maximum intensity to 30. The signal-to-noise ratio (SNR), after reconstruction by different methods, is also given. c, Pearson correlations of results obtained by VsLFM, VCD-Net and HyLFM-Net, compared with ground truth. The center line represents the median, the box limits represent the lower and upper quartiles, and the whiskers represent 1.5-fold the interquartile range. P values were calculated using the two-sided paired t-test: P = 9.40 × 10⁻¹⁸ for VCD-Net and P = 8.77 × 10⁻¹⁷ for HyLFM-Net. n = 17 for each method, which represents the number of noisy images. d, Reconstructed MIPs with an induced aberration wavefront, the root mean square (RMS) of which was set to 1 wavelength, obtained by VsLFM without DAO, VsLFM with DAO, VCD-Net and HyLFM-Net. The estimated wavefront by DAO is shown in the inset. The Fourier transforms corresponding to the whole FOVs by the four methods are shown in the right panel. e, Normalized intensity profiles along the blue dashed line marked by the arrows in d for four different methods. The arrows indicate the positions of the signal peak. f, The curves of reconstructed SSIM versus aberration levels applied for different methods. Note that Vs-Net, VCD-Net and HyLFM-Net used here were all trained on the same tubulin data in ideal imaging conditions without noise and aberration. Scale bars, 10 μm (a,b), 10 μm (left) and 2 μm⁻¹ (right) (d).