Large-scale microscope with improved resolution using SRGAN

Abstract

Due to the limited Space-Bandwidth Product (SBP) of optical microscopy system, the Numerical Aperture (NA) and Field of View (FOV)/ Depth of Field (DOF) of conventional optical microscopes are mutually restricted. In this paper, we propose a 20× large-scale (large FOV&DOF) microscope objective lens which is enhanced by deep neural network to achieve the improved resolution close to a 40× Olympus objective lens. Based on training a Super-resolution Generative Adversarial Network (SRGAN) model to transform low-resolution images into super-resolved ones, the resolution of input images captured by low magnification objective is greatly improved without any loss of FOV. And the network output images have better performance on DOF. We also compared SRGAN with two other deep-learning based methods to prove its superiority. Such large-scale microscope with improved resolution could serve to democratize super-resolution imaging.