Untrained neural network enabling fast and universal structured-illumination microscopy

Abstract

Structured-illumination microscopy (SIM) offers a twofold resolution enhancement beyond the optical diffraction limit. At present, SIM requires several raw structured-illumination (SI) frames to reconstruct a super-resolution (SR) image, especially the time-consuming reconstruction of speckle SIM, which requires hundreds of SI frames. Considering this, we herein propose an untrained structured-illumination reconstruction neural network (USRNN) with known illumination patterns to reduce the amount of raw data that is required for speckle SIM reconstruction by 20 times and thus improve its temporal resolution. Benefiting from the unsupervised optimizing strategy and CNNs’ structure priors, the high-frequency information is obtained from the network without the requirement of datasets; as a result, a high-fidelity SR image with approximately twofold resolution enhancement can be reconstructed using five frames or less. Experiments on reconstructing non-biological and biological samples demonstrate the high-speed and high-universality capabilities of our method.