Classification Of Cellular Diffraction Images Based On Deep Transfer Learning

As basic units of life,cells’ structures are fundamental to their functions.Research of new methods for accurate analysis and classification of single cells is of great significance in the fields of cell biology,clinical diagnosis and treatment of disease.At present,conventional methods to analyze single cells such as conventional flow cytometry and immunofluorescence microscopy require fluorescent labeling of cells,and they may be either unable to provide detailed morphological information or time-consuming and costly.To solve the above problems,we built polarization diffraction imaging flow cytometry(p-DIFC)and collected the polarization diffraction images that can reflect the morphological structure of cells.We designed a convolutional neural network for polarization separation of diffraction images,proposed a classification method combining contrastive learning and migration learning,and conducted a label-free classification application study on human lymphocytes.To address the problem of large variability of light intensity information in different polarization directions in cell diffraction images,we propose a new method for cell classification using Separated Polarization Convolutional Neural Networks(SP-CNN)in this report,which can effectively extract the cell diffraction features in different polarization directions by using two different network structures for orthogonal polarization diffraction images and achieves label-free and accurate classification of five types of cultured cell lines.The average classification accuracy was found to be 99.2%.It achieves label-free classification of cells than traditional cell classification methods and has higher classification accuracy than other classification networks.Aiming at the problems of insufficient labeled data and time-consuming labeling of diffracted images,we propose a new Contrastive Learning of Cell Polarization images pretraining networks(CLCP)in this report.The method combines the knowledge of cell pairing and polarization pairing to efficiently extract cell diffraction image features without using additional label information,and a pre-trained model is trained.The downstream classification experiments based on the pre-trained model with migration learning achieve 93.8% classification accuracy with only 20 labeled images per cell class,providing an efficient and accurate classification method for cell diffraction images with few samples.Aiming at the problem of inaccurate antigen labeling when antibodies from human lymphocytes react with similar strength to different antigens,we propose a human lymphocyte classification method based on deep learning and peak density clustering algorithm in this thesis.The method selects a set of support cells by the Pearson correlation coefficient between cell features and divides all the cells into two groups A and B by using them as criteria,among which group A accounts for 90.2%of the total,and the classification accuracy of group A using SP-CNN is 97.3%.Therefore,in this paper,group A cells are regarded as correctly labeled cells and group B cells are regarded as incorrectly labeled cells.Finally,this paper further demonstrates the effectiveness of the method for cell labeling detection by depolarization and feature visualization analysis,which has important biological research and medical applications.