Fig. 1. Schematics of deep learning architecture. (a) U-Net network architecture^[97]; (b) schematic of CGAN^[105]; (c) schematic of RCAN^[112]; (d) schematic of FCA^[110]

Fig. 2. Confocal microscopy results based on deep learning. (a) Restored images using the UTOM algorithm^[106], where the scale bar represents 50 μm; (b) super-resolution imaging using BPGAN^[107], where the scale bars are 10 μm and 20 μm; (c) imaging on neuronal mitochondria using PSSR^[99], where the scale bar represents 10 μm

Fig. 3. Imaging results of deep learning in light-sheet fluorescence microscopy. (a) Reconstruction result of zebrafish hearts using DVSR algorithm^[91]; (b) neural/nuclear reconstruction of the Deep-SLAM algorithm in mouse brain^[100], where the scale bar is 100 μm; (c) fluorescent microspheres/mouse brain slice images obtained using CBS-Deep algorithm^[90]

Fig. 4. Schematic of light field fluorescence microscope^[120]

Fig. 5. Imaging results of deep learning in light-sheet fluorescence microscopy. (a) Imaging results of various cardiac dynamics in beating zebrafish heart using VCD-LFM^[101], where the scale bar is 50 μm; (b) imaging results of HyLFM algorithm on medaka heart^[92], where the scale bar is 50 μm

Fig. 6. Structure and principle of SIM system^[122]

Fig. 7. SIM imaging results based on deep learning.(a) Cross-modal imaging results from TIRF to TIRF-SIM^[108]; (b) U-Net-SIM3 super-resolution imaging effect under extreme low-light conditions^[102], where the scale bar is 1 μm; (c) 3_SIM imaging results^[93]; (d) 5_NSIM imaging results^[109]; (e) imaging results of DFCAN/DFGAN^[110], where the scale bar is 2 μm

Fig. 8. Principle of STED microscopy imaging^[122]

Fig. 9. STED imaging results based on deep learning. (a) Super-resolution imaging results using SRDAN^[111]; (b) multimodal imaging results from confocal images to STED super-resolution images^[108]. Scale bar of first row: 2000 nm. Scale bar of second row: 300 nm

Fig. 10. Improved imaging performance of single molecular positioning microscope by the deep learning algorithm. (a) Reconstructed microtubule images using Deep-STORM^[95], where the scale bar is 0.5 μm; (b) super-resolution three-dimensional imaging using DeepSTORM3D^[96], where the scale bar is 5 μm; (c) volume tracking results of telomeres in living mouse embryonic fibroblast (MEF) cells using DeepSTORM3D^[96], where the scale bar is 2 μm

Fig. 11. Deep learning cloud computing platform ZeroCostDL4Mic ^[141].(a) Schematic of ZeroCostDL4Mic; (b) workflow of ZeroCostDL4Mic; (c) bioimage analysis tasks implemented with the ZeroCostDL4Mic platform