Resolution enhancement in microscopic imaging based on generative adversarial network with unpaired data

Abstract

Deep learning has shown the high potential for the use of resolution enhancement in microscopic imaging. However, a large amount of paired datasets are usually required for training the networks. Data preparation is costly from the microscopic experiment. In this paper, the cycle generative adversarial network, trained by only a small amount of unpaired experimental microscopic images, is used to generate paired data for supervising learning in resolution enhancement. A typical network for resolution enhancement, named Pix2Pix-GAN, is then trained by the generated images to generate high-resolution images. A various of samples have been tested. From the images obtained with low-power microscopes, the trained network can output the images similar to those acquired with high-power microscopes. Results show that the generated dataset can effectively replace the experimental one for neural network training. Problems related to the experimental paired data preparation are then solved. The proposed method further broadens the application of deep learning for the optical field, especially for resolution enhancement in dynamic microscopic imaging.