Extended Data Fig. 7: Temporally and Spatially interleaved self-supervised learning of rDL denoising networks.
a, Representative LLSM images of Golgi SiT-vesicles denoised by TiS-rDL self-supervised and rDL supervised learning models. Depth color-coded maximum intensity projections (MIPs) of input noisy raw data (left), the images stack denoised by the TiS-rDL model (middle) and by the supervised learning model (right). These results illustrate the performance of TiS-rDL denoising model is as good as the rDL model trained with the supervision of high-SNR images. b, c, Representative TiS-rDL denoised images of Golgi SiT-vesicles (b) and statistical analysis in terms of PSNR under the conditions of temporally down-sampled at different factors (c), n = 50. Center line, medians; limits, 75% and 25%; whiskers, maximum and minimum. Noisy raw images and images denoised by the supervised model are shown for reference. These results indicate that the decrease of temporal sampling rate gradually affects the output fidelity of the TiS-rDL method. d, e, Schematic of SiS-rDL denoising network training (d) and inference (e). f, g, Representative outputs (f) and PSNR comparisons (g) of SiS-rDL denoising models with and without gap-amending regularization (GAR) in terms of ER and Mito in mitotic cells at different axial sampling intervals of 50, 100, 200, 300, 400, 500, and 600 nm, n = 31 for ER, 42 for Mito. Center line, medians; limits, 75% and 25%; whiskers, 95% and 5%. These results show that GAR effectively improves the performance of SiS-rDL models; The PSNR of SiS-rDL denoised images decreases by less than 4% before the axial sampling rate is lower than Nyquist criterium, that is, 300 nm for LLSM, which suggests the general applicability of the SiS-rDL method for 3D imaging. Scale bar, 2 μm (a, b), 5 μm (f), 3 μm (zoom-in regions of f).
