Extended Data Fig. 14: Simulation of the deep-learning method for isotropic in-plane super-resolution imaging.

a) An object (left column) consisting of lines and hollow spheres can be blurred to resemble diffraction-limited confocal input (middle column). Fourier transform of the raw input is shown (right column). b) Deep-learning 1D super-resolution output with no rotation (left column); deep learning output after rotating input by 30 degrees (middle column); deep learning output after rotating input by 60 degrees (right column). Fourier transforms in bottom row confirm 1D resolution enhancement regardless of rotation angle. c) Deep learning output after rotating input by 90 degrees (left column); outputs after deep learning and joint deconvolution of two orientations (middle column), or six orientations (right column) show progressive improvement in resolution isotropy (red arrows), confirmed with Fourier transforms of the images (bottom row). Ellipses bound decorrelation estimates of resolution (numerical values from ellipse boundary indicated in red text). Scale bar: 2 μm.