Fig. 1: Self-Net pipeline and validation of its isotropic recovery performance using synthetic and semisynthetic data.

a Self-learning strategy and schematic of Self-Net. b First row: XZ maximum-intensity projections (MIPs) of the raw anisotropic data, isotropic ground truth (GT) data, OT-CycleGAN, and Self-Net restorations of simulated fluorescence beads. The projection thickness is 50 μm. Second row: corresponding color-merged images by merging the data (magenta) and GT (green). The insets show enlarged views of the corresponding dashed boxes. Scale bar: 5 μm; 1 μm for the insets. c Intensity profiles along the corresponding colored lines in the insets in (b). d First row: XZ MIPs of the raw anisotropic data, isotropic GT data, OT-CycleGAN, and Self-Net restorations of synthetic tubular volume. The projection thickness is 30 slices. The yellow arrowheads highlight the details inconsistent with the GT data in the OT-CycleGAN output. Second row: corresponding color-merged images by merging the data (magenta) and GT (green). Scale bar: 20 pixels. e Quantitative comparison of the image qualities of d via the RMSE and the SSIM. Results are obtained from five randomly-selected MIP images of 900 × 900 pixels. f Comparison of model size, training time, and inference time between the OT-CycleGAN and Self-Net using the same GPU card (GeForce RTX 3090, Nvidia). The inference time was evaluated using a 700³ voxels volume. g MIPs of the raw image stack (top right: PSF kernel used to generate anisotropic blurring), GT image stack, CARE restoration (supervised learning), and Self-Net restoration (unsupervised learning) of semisynthetic neuronal data. The projection thickness is 300 μm. Normalized intensity profiles along the corresponding white dashed lines in the enlarged images are shown in the bottom-left corner of the images. Scale bar: 10 μm; 5 μm for the enlarged images. h Quantitative comparison of the image qualities of g via RMSE and SSIM (n = 5 independent image stacks)