Fig. 1. Principle of SelfMirror and visualization of denoising three-dimensional imaging data. (A) Self-supervised learning strategy of SelfMirror. An imaging z stack was obtained by volumetric imaging. The original noisy image sequence is split into two sub-sequences. Each sub-stack is fed into a separate but identical network for training. The parameters of the two networks are shared. (B) The inference scheme of the trained SelfMirror; the whole noisy stacks would be fed into the network. (C) An example of 3D volume rendering of the neuronal population. The left shows the raw low-SNR volume, the middle shows the SelfMirror denoised volume, and the right shows the GT reference volume. Scale bars, 50 μm. (D) Magnified views of neuronal soma and spine structures in the color-boxed regions in (B). Scale bars, 15 μm.

Fig. 2. Performance validation on simulated data. (A) Max projection along three (z, y, x) axes of blue-boxed region in Fig. 1A. From left to right are the GT, noisy, SelfMirror (ours), DIP, N2V, N2F, N2S, Ne2Ne, and SUPPORT denoised data. The images are colored with red hot. Scale bars, 20 μm. (B) From left to right are the pixel intensity profiles along the white dashed lines of max-z, max-y, and max-x in (A), respectively. The intensity is normalized to 0–1.

Fig. 3. SelfMirror denoises two-photon volumetric imaging data of single neurons. (A) Visualization of single neuronal morphological images with a 340μm×340μm×220μm volume (278 planes, 0.66 μm/pixel in x and y, 0.8 μm/pixel in z) in the mouse cortex. From left to right are the original noisy volume, the same volume denoised with SelfMirror, and the high-SNR reference volume. Color-boxed regions show the magnified views of the neuron structures with soma, dendrite, and spine. Scale bars, 50 μm for the whole FOV and 20 μm for magnified views. (B) Orthogonal views of the same imaging plane of the neuronal volumes. From left to right are the noisy data, SelfMirror denoised counterparts, and high-SNR reference data. Magnified views of blue-boxed regions are shown at the bottom right of the images. Scale bars, 50 μm for the whole FOV and 15 μm for magnified views. (C) From left to right are the pixel intensity profiles along the blue-, white-, and red-dashed lines in (B), respectively. The intensity is normalized to 0–1. (D) Left, the statistical evaluation of Pearson correlation coefficient increases of traces before and after denoising by SelfMirror, in which the trace is randomly selected in xy slices, N=32. Each line represents one of 32 traces, and increased correlations are colored blue. Right, Tukey box-and-whisker plot of the fluorescent traces. p values calculated by two-sided one-way paired t-test are specified with asterisks, ***p<0.001. (E) Left, the statistical evaluation of Pearson correlation coefficient increases of traces before and after denoising by SelfMirror, in which the trace is randomly selected in xz and yz slices, N=31. Each line represents one of 31 traces, and increased correlations are colored blue, decreased correlations are colored red. Right, Tukey box-and-whisker plot of the fluorescent traces. p values calculated by two-sided one-way paired t-test are specified with asterisks, ****p<0.0001.

Fig. 4. SelfMirror denoises two-photon volumetric imaging data of neuronal population. (A) Visualization of morphological structure imaging of neuronal population with a 340μm×340μm×220μm volume (144 planes, 0.66 μm/pixel in x and y, 1.5 μm/pixel in z) in the mouse cortex. From left to right are the original noisy volume, the same volume denoised with SelfMirror, and the high-SNR reference volume. Scale bars, 50 μm. (B) Magnified views of the neuron structures with sliced soma, and dendrite in the three color-boxed regions at three different depths (52,96,142 μm) in (A). From left to right are the original noisy data, the same data denoised with SelfMirror, and high-SNR reference data. Scale bars, 20 μm. (C) Example orthogonal frames of the neuronal volumes. Left, original noisy frames from three-dimensional views. Middle, SelfMirror denoised counterparts. Right, high-SNR reference frames. The imaging planes are the same on the left, middle, and right. A magnified view of the red-boxed region is shown at the bottom right of the images. Scale bars, 50 μm for the whole FOV and 10 μm for magnified views. (D) Pixel intensity profiles along the blue-dashed lines in (C). The intensity is normalized to 0–1. (E) Statistical spectrum plots of the intensity value of the blue-boxed signal-free region in (C), along the imaging z-axis. One row of the spectrum plot represents all pixels (n1=2800) of the blue-boxed region in (C). The column direction of a spectrum plot represents the imaging planes (n2=144). The equation for total pixels of a plot is n=n1×n2=403,200. All pixels were normalized with a plasma bar. Zoom-in blue-boxed region views are shown in the right panel. (F) Representative z-axis slice of structural imaging of dendrite and spine with a 180μm×180μm×150μm volume (100 planes, 0.35 μm/pixel in x and y, 1.5 μm/pixel in z) in a mouse expressing GFP. Left, the original imaging data without AO as low-SNR image. Middle, the same image denoised with SelfMirror. Right, imaging data with AO as a high-SNR reference. Magnified views of the blue-boxed region at multiple axial locations are shown in the bottom panel. Axial location of 35 μm corresponds to the current frame. Scale bars, 25 μm for the whole FOV and 10 μm for magnified views.

Fig. 5. Denoising structural imaging data from multiple fluorescent microscopies. (A) Visualization of representative z-axis slices (top) and y-axis slices (bottom) from left, noisy data, middle, corresponding SelfMirror denoised data, and right, high-SNR data of Penicillium. The pixel intensity profiles of the blue- and white-dashed lines are inserted on the left bottom of the images, respectively. Scale bars, 20 μm. (B) Magnified views of the blue- and yellow-boxed regions in (A) at multiple axial locations. Scale bars, 2 μm for all images. (C) Box-and-whisker plot showing PSNR and Pearson correlation coefficient for z-axial slices before and after SelfMirror denoising; high-SNR reference data were used as the ground truth for calculation. A two-sided paired-sample t-test is used, N=129, which represents the number of planes along the z-axis (****p< 0.0001). (D) STD projection (top) and max projection (bottom) of vessels of a mouse cortex for the raw noisy data (left) and corresponding images using SelfMirror (right). Scale bars, 30 μm. (E) Example frames in three-axis planes (top, z-axis; bottom left, y-axis; bottom right, x-axis) of the intestine of a mouse embryo after expansion for the original noisy data (left) and corresponding denoised image using SelfMirror (right). Scale bars, 30 μm.

Fig. 6. Denoising multiple volumes from multiple imaging modalities. (A) Representative slice from low-SNR (left), SelfMirror denoised (middle), high-SNR (right) CT volumes of human thoracoabdominal body. (B) Low-SNR (left) and SelfMirror denoised (right) error maps with high-SNR image in (A). Error maps are colored with a plasma bar. (C) Box-and-whisker plot showing Pearson correlation coefficient (left) and structural similarity index (right) of axial slices. A paired-sample t-test is used, N=560, which represents the number of planes along the z-axis (****p< 0.0001). (D) Example xy plane (left) and yz plane (right) frames of a mouse somatosensory cortex layer 4 from a 3D-EM volume 93μm×60μm×93μm. Scale bars, 10 μm.