Reconstruction of structured illumination microscopy with an untrained neural network

Structured illumination microscopy (SIM) is one of super-resolution optical microscopic techniques, and it has been widely used in biological research. In this paper, a physics-driven deep image prior framework for super-resolution reconstruction of SIM (entitled DIP-SIM) is proposed. DIP-SIM does not rely on a large number of labeled data, and the output becomes more interpretable due to the intrinsic constraint of a physical model. Both the simulation and experiment verify that DIP-SIM can reconstruct a super-resolution image with a quality comparable to conventional SIM. Of note, it allows for super-resolution reconstruction from three raw images for two-orientation SIM and four raw images for three-orientation SIM, and hence it has a much faster imaging speed and lower photobleaching compared with the traditional SIM. We can envisage that the proposed method can be applied to chemistry and biomedical fields, etc.