Extended Data Fig. 5: Comparison of RL-DFN with existing axial resolution enhancement methods.
a-d, Representative x-y slices and x-z slices of simulated spherical shells (a), simulated tubular structures (b), experimentally acquired MTs (c), and experimentally acquired ER (d), reconstructed by multiple axial resolution enhancement methods, including ID-Net19, DL-ARE20, Self-Net21, multi-view RL deconvolution27, and the proposed RL-DFN. Anisotropic inputs, 3D rendering of the inputs, and the GT images are shown for reference. f,g, Statistical comparisons of ID-Net, DL-ARE, Self-Net, multi-view RL deconvolution, and RL-DFN in terms of PSNR and SSIM (n = 100) using simulated data of spherical shells (f) and tubular structures (g). Central line, medians; limits, 75% and 25%; whiskers, maximum and minimum. Scale bar, 2 μm (3D rendering images in a-d), and 1 μm (slice-images in a-d).
