Autofocus Phase Imaging Technology Based On Deep Learning

Cell organization is an important part of an organism.By analyzing cell structure,we can understand the formation process of cells and study their internal structural characteristics.But most cells are phase objects that cannot be directly imaged by conventional microscopy.The phase-contrast method provides a microscope technique to observe samples without labeling and staining,but Zernike phase-contrast microscopy requires a circular Zernike phase plate embedded in the Fourier spectrum of the objective lens.Differential interference phase-contrast microscopy requires a birefringent crystal.The complex optical configuration and computational time cost limit the further application of its phase contrast method,and the phase contrast method cannot quantitatively measure the object phase information.With the development of technology,many quantitative phase imaging methods have been proposed.Digital holographic microscopy is a three-dimensional measurement method with high accuracy,non-contact and non-damage.However,it has some limitations in phase unwrapping,resolution,image signal-to-noise ratio and other aspects,which limit the further application of this technology.In addition,autofocusing is a key step to obtain high-quality phase information,and autofocusing and phase imaging must be achieved simultaneously.This thesis focuses on the above problems.Aiming at the shortcomings of traditional phase contrast imaging methods and the key of autofocus to image quality,a deep learning method was proposed to achieve autofocus and phase contrast imaging at the same time,and the feasibility of the method was preliminarily verified by Pix2 pix and U-Net.Aiming at the characteristics of autofocus phase contrast imaging and the shortcomings of Pix2 pix and U-Net,a network framework based on U-Net was designed by introducing spatial attention mechanism and residual block.The test results and comparative analysis show that the framework can achieve autofocus phase contrast imaging more accurately.In addition,the effect of different FOV on this network was tested.Finally,we discuss how to achieve autofocus phase contrast imaging in the absence of matching data sets.At present,no microscope can directly obtain the focused phase map from the defocus micrographs under incoherent illumination.To solve this problem,a deep learning method was proposed to achieve cross-modal auto-focus quantitative phase imaging,and the method was capable of unsupervised(unpaired image data)learning.In this thesis,the effectiveness of the proposed strategy was first qualitatively verified using defocus microscopic images and mismatched phase maps.In addition,the accuracy of the output images of the network was verified by quantitative testing of one-to-one corresponding intensity and phase maps.