Extended Data Fig. 1: Illustration of the spectral bias effect in deep learning super-resolution (DLSR) task.
a, Concepts (upper row) and typical results (lower row) of wide-field, DFCAN-SIM, rDL SIM and conventional SIM imaging. b, Typical DFCAN-SIM images of F-actin output at 4000, 24000, 52000, 100000, 200000, and 400,000 training iterations, respectively. c, The evolutions of the training loss and the resolution of the intermediate outputs during the training process of the DFCAN-SIM model of F-actin dataset (averaged from n = 20 test images). The curves shows that there exists a significant resolution gap between DFCAN-SIM (red dashed line) and GT-SIM (black dashed line) images, which implies DLSR models suffer from severe spectral bias. d, The corresponding GT-SIM image of the same ROI with (b) for reference. e, Typical rDL denoised images (top left) and rDL SIM images (bottom right) of F-actin output at 4000, 24000, 52000, 100000, 200000, and 400,000 training iterations, respectively. f, The evolutions of the training loss and the resolution of the intermediate outputs during the training process of the rDL denoising model of F-actin dataset (averaged from n = 180 test images). Compared with DLSR task in b and c, the rDL denoising of raw SIM images is hardly impaired by the spectral bias problem, preventing the resolution degradation of rDL SIM image relative to the GT-SIM image, as illustrated in (a). g, The corresponding ground truth raw SIM image of the same ROI with (e) for reference. Scale bar, 3 μm (a, b, d, e, g), 0.5 μm (zoom-in regions of a, b, d, e, g). Gamma value, 0.7 for DFCAN-SIM images, rDL SIM images and GT-SIM images.
