Fig. 1. Noise model of super-resolution fluorescence microscopy

Fig. 2. Structure of proposed convolutional neural network

Fig. 3. Simulated noise picture and noise-free picture. (a) Simulated low signal-to-noise ratio wide-field fluorescence signals with 500 photons of background and 400 photons of signal; (b) noise-free wide-field fluorescence signal

Fig. 4. Comparison of noise reduction capabilities of different algorithms on simulated fluorescence images. (a) Simulated fluorescence image with low signal-to-noise ratio; (b) fluorescence image after noise reduction by Gaussian filter algorithm; (c) fluorescence image after noise reduction by BM3D algorithm; (d) fluorescence image after noise reduction by proposed convolutional neural network algorithm; (e) noise-free fluorescence image

Fig. 5. Peak signal-to-noise ratio of fluorescence images recovered by different algorithms.(a) Peak signal-to-noise ratio of fluorescence images recovered by different algorithms when the number of signal photons gradually changes from 100 to 900 under the condition that the number of background photons is 500; (b) peak signal-to-noise ratio of fluorescence images recovered by different algorithms when the number of background photons gradually changes from 100 to 900 under the condition that the numbe

Fig. 6. Schematic of the reconstruction method using proposed convolutional neural network and Deep-STORM

Fig. 7. Simulated imaging target and reconstruction image. (a) Original image obtained by simulation; (b) diffraction-limited wide-field image; (c) Deep-STORM reconstructed super-resolution image under noise-free condition

Fig. 8. Fluorescence images recovered by different noise reduction algorithms. (a) Simulated fluorescence images with low signal-to-noise ratio; (b) simulated noise-free fluorescence image; (c) fluorescence image recovered by proposed convolution neural network algorithm; (d) fluorescence image restored by BM3D algorithm; (e) fluorescence image recovered by Gaussian filter algorithm

Fig. 9. Reconstructed super-resolution images by different denoise algorithms and relative intensity distribution. (a) Reconstructed super-resolution image after noise reduction with proposed convolutional neural network algorithm; (b) reconstructed super-resolution image after noise reduction with Gaussian filter algorithm; (c) reconstructed super-resolution image after noise reduction with BM3D algorithm; (d) directly reconstructed super-resolution image ; (e)--(h) relative intensity

Fig. 10. Low signal-to-noise ratio fluorescence image and wide field image of endoplasmic reticulum collected in experiment. (a) Fluorescence image; (b) wide-field image of endoplasmic reticulum

Fig. 11. Fluorescence images denoised by different algorithms and reconstructed super-resolution images after noise reduction. (a) Restoration of fluorescence image obtained by proposed convolutional neural network; (b) restoration of fluorescence image obtained by BM3D algorithm; (c) restoration of fluorescence image obtained by Gaussian filter algorithm; (d) reconstructed super-resolution image after noise reduction with proposed convolutional neural network algorithm; (e) reconstructed super-resolutio

Fig. 12. Fluorescence image with low signal-to-noise and wide field images of microtubules. (a) Fluorescence image with low signal-to-noise; (b) wide-field image of microtubules

Fig. 13. Fluorescence images of microtubules after noise reduction by different algorithms and their super-resolution images. (a) Fluorescence images after noise reduction by different algorithms; (b) super-resolution images reconstructed after noise reduction by different algorithms; (c) enlarged microtubules