Time-series Image Classification Of Cells Based On Deep Learning

In medical research,it is usually necessary to achieve some research purposes according to the morphology and survival rate of cells or organoids.There is a common problem in cell research,that is,time-series image classification of cells.It is often necessary to extract some temporal and spatial information from the images of cells changing over time to classify the time-series images of cells.Based on the different data forms of cell images,this thesis divides the classification of cell time-series images into the following three categories: 1.Each image contains a chronological image of multiple cells;2.Each image contains a cell timing image of a cell subject;3.A time-series image of cells changing over a period of time is captured in the form of video.Combined with the method of deep learning,this thesis proposes solutions for the above three kinds of cell time-series image classification problems and improves the algorithm.In addition,three different scenario tasks were presented to solve the above three problems:time-series image classification of umbilical cord mesenchymal stem cells by culturing multiple generations of umbilical cord mesenchymal stem cells and taking cell images.Time-series images of cancer cells under drug action were analyzed to determine the type of drug.The type of drug was determined according to the myocardial sphere video before and after drug action.The main research contents are as follows:1.The identification of umbilical cord mesenchymal stem cells was studied in this thesis.We took several cell pictures of each generation of several umbilical cord mesenchymal stem cells cultured for several generations,and identified the umbilical cord mesenchymal stem cells by cell morphology and the differences between each generation.In this thesis,an improvement was made on the Res Net-LSTM network.Based on the multi-scale and attention mechanism,the MSA-Res Net-LSTM network was proposed.The accuracy rate of94.8% was achieved in the test set,and most cell line images could be correctly recognized.2.The temporal image classification of cancer cells under the action of drugs was studied in this thesis.In this thesis,the time-series images of cells under the action of drugs are used.Based on the method of deep learning,I3 D network in 3D convolutional neural network is taken as the basis,and a funnel convolutional structure that can independently extract time and space features is proposed to replace 3D convolutional kernel,thus improving the network.The accuracy of 91.4% was obtained in the test set,and it can be used for drug screening.3.The experimenter wanted to know the type of drug by watching the video of the myocardium contracting.In this thesis,image segmentation and frame difference method are used to analyze the video of myocardial sphere contraction,and segmented multi-scale entropy algorithm and MSEAMP index are proposed to classify the curves.This index can be used to classify cisapride,nifedipine and isoproterenol.