Research On Quantitative Phase Imaging Based On Generative Adversarial Network

With the rapid development of modern life science,people's cognition and research on microscopic life activities are becoming more and more in-depth.In order to better observe biological tissues,cells,molecules,and other microscopic biological samples,researchers have become increasingly demanding imaging techniques with high resolution,fast detection speed,and no damage to biological samples,which has also accelerated the research on quantitative phase imaging technology combining artificial intelligence and imaging technology.In recent years,the rapid development of artificial intelligence has provided new ideas and technical means for quantitative phase imaging.The phase reconstruction technology based on deep learning can quickly and accurately reconstruct the phase from a hologram to achieve dynamic phase reconstruction,avoiding the introduction of mechanical errors and unnecessary noise.However,in practical applications,problems such as difficulty in obtaining data sets and poor generalization ability of deep learning networks limit the further application of deep learning in quantitative phase imaging.In response to the above problems,this paper proposes a data-driven quantitative phase imaging method,which uses a phase imaging-oriented generative adversarial network PI-GAN(Phase Imaging Generative Adversarial Network)to train from largescale high-quality simulation data sets.And it uses transfer learning to achieve the balance of generalization ability and precision,and finally achieve fast and accurate quantitative phase imaging of biological samples from an interference fringe pattern.It provides important technical reference for realizing quantitative phase imaging based on deep learning.The main research content and research results obtained include the following aspects:1.This paper studies deep learning algorithms for quantitative phase imaging,and proposes the idea of using simulated interferogram training,verification on experimental interferogram,and improving generalization ability through transfer learning.This solves the problem of difficulty in obtaining the data set,improves the diversity of the data set and the quality of the label,and also significantly enhances the generalization ability of the neural network.2.This paper studies the principle of coaxial holographic phase imaging.A data set production method is proposed,which simulates phase objects based on the principle of Fourier series expansion,and extracts a priori parameters from the experimental interferogram to improve the quality of the simulated interference fringe pattern.The production method finally obtains large-scale and high-quality simulation data sets and uses them for the training of neural networks.3.In this paper,we propose a PI-GAN architecture for quantitative phase imaging,and build the network,train it with simulation data sets,and test and evaluate it on the experimental interferogram.The PI-GAN is compared with the traditional quantitative phase imaging method.The experimental results show that PI-GAN can quickly and accurately perform quantitative phase imaging of biological samples from an interference fringe pattern.