Research On The Method Of Phase Object Recognition And Analysis Based On Quantitative Phase Imaging And Deep Learning

A biological cell is the most basic structural and functional unit of life,the recognition and analysis of their morphological structures and other information have great significance in the fields of life science and clinical medicine.Most cells are colorless and transparent,but belong to phase objects,and it is difficult to image cells with traditional microscope.Quantitative phase imaging techniques can modulate the invisible phase information of the light wave into amplitude information which is easy to observe and extract.Thus the static morphological structure information or dynamic changes of a phase object can be achieved quantitatively with such a noninvasive label-free detection approach.By applying this technique to cell recognition and disease diagnosis,the analysis efficiency can be improved and the subjective errors can be reduced effectively.But for samples with special shapes or complicated internal structures,it is difficult to extract the subsequent quantitative parameters as the refractive index distribution of the sample and the physical thickness are coupled in the phase shift information.Besides,there are situations in which the parameters manually set are not powerful enough to describe the unknown sample.Therefore,the research on automatic recognition and analysis of quantitative phase information of samples needs to be further deepened.This paper proposes a method to recognize and classify different phase objects based on the different phase distribution of samples using deep learning technology.Phase images of four different types of samples with similar contours were collected,and the classic convolutional neural network model was optimized for the recognition.On this basis,the sample data is manually transformed to further narrow the differences between samples,and the experimental results show that the trained convolutional neural network has a good ability to recognize the phase distribution images of different samples.In order to simplify the recognition process,this paper further explores the method of recognizing the interferograms collected during the process of phaseimaging to classify different phase objects,thereby avoiding the tedious phase numerical reconstruction and the errors it may introduce.Then we introduced the residual block and tested the ability of the convolutional neural network and the residual network to recognize the interference fringes.By changing the spatial frequency and visibility of the fringes,the performance of the network model for the recognition of interferograms with different modulation factors is analyzed through comparative experiments.Furthermore,a deep learning strategy for efficiently recognizing interferograms is summarized,and the feasibility and accuracy of the strategy are preliminarily proved through experiments.This paper also introduces a phase imaging system integrated with optical scattering information to solve the limitations of the phase recognition based on deep learning.The system can obtain the phase information and scattering information of the sample in batches at the same time.The simulation and experiment show that the scattering signal can be used to qualitatively analyze the internal structure of the sample and thus assist the phase analysis to extract the morphological characteristics and substructure location information of the sample.