Fig. 1. Neural network model framework. (a) Multilayer perceptron; (b) convolutional neural network (CNN) based feature network; (c) Inception module; (d) residual network (ResNet); (e) encoder-decoder architecture; (f) long short-term memory (LSTM) network; (g) Transformer; (h) generative adversarial network

Fig. 2. Single-molecule localization method DeepSTORM based on deep learning and simulated dense data^[9]. (a) Network architecture of DeepSTORM; (b) diffraction-limited low resolution image (left), DeepSTORM reconstruction with ground truth emitter positions (red crosses) (middle), and magnified view of the selected region (right)

Fig. 3. Image reconstruction processes of DI-STORM and conventional SMLM method^[85]

Fig. 4. Design of an optimial PSF using neural networks for multicolor imaging^[80]. (a) Amount of phase change of each wavelength after passing through SLMs with same voltage; (b) optimization flow of multi-wavelength PSF based on neural network; (c) optimized SLM voltage pattern for color determination by neural network; (d) phase delay for 565 nm, 625 nm, 705 nm and 800 nm light with same voltage; (e) simulated PSFs of different wavelengths

Fig. 5. Workflow of machine learning-based multidimensional SMLM^[10]. (a) Color-separating ANN; (b) ANNs for resolving axial position; (c) analysis of unknown samples

Fig. 6. Data simulation based on SMLM simulator software. (a) Three-dimensional simulation of microtubules by SuReSim^[97]; (b) TestSTORM simulated axonal patterns (left) and generated TIFF image (right)^[98]; (c) SMeagol simulated 　　　　fluorescently labeled MinE proteins^[99]