Deep learning for fast image reconstruction of Fourier ptychographic microscopy with expanded frequency spectrum

Abstract

Fourier Ptychographic Microscopy (FPM) is a super-resolution microscopy technology, in which a set of low-resolution images containing different frequency components of the sample can be obtained by changing the angle of the light source in this technology, and then the iterative algorithm is used to reconstruct high-resolution intensity and phase information. The reconstruction usually takes a long time and is not suitable for real-time FPM imaging. It has been recognized recently that the potential fast image reconstruction algorithm is the use of deep learning algorithms. We designed a conditional generative adversarial network (cGAN) which has multi-branch input and multi-branch output which can expand the frequency spectrum of the reconstructed image very well. Based on the convolutional neural network (CNN), the brightfield and darkfield images obtained by FPM imaging can be regarded as different image features obtained by different convolutional kernel, and the skip connection of U-net can effectively utilize this information. The brightfield and darkfield images in FPM imaging are input to different branches, which can avoid missing the darkfield signal information. Importantly, the neural network we designed will continue to perform simulation process of FPM imaging from the recovered high-resolution intensity and phase to obtain low-resolution images and make them correspond one-to-one with the input low-resolution images. These corresponded images will enter loss function, making it easier for the neural network to learn relation between the low-resolution images and the high-resolution images. We validated the deep learning algorithm through simulated experimental research on biological cell imaging.