Research On The Diffraction Imaging Classification Of Label-free Human Prostate Cells

The structural changes of human cells,especially the changes in the size of the nucleus can provide an important evidence for cancer diagnosis.At present,histological analysis and fluorescence microscopy are still the preferred methods for cell detection.In recent years,cellular fluorescence microscopy has made progress,but these techniques are based on incoherent imaging,which requires well-trained experts to stain,collect and read.They are time-consuming,expensive,and staining will damage the cells,which is a destructive inspection.Therefore,there is urgent to find a non-destructive testing method without staining.In this paper,the single cell of prostate was stimulated by coherent light,and the polarization diffraction imaging was realized.The feature parameters of the polarization diffraction image(pDI)were extracted by a Gray-Level Co-occurrence Matrix(GLCM),and normal and cancerous prostate cells were accurately classified based on machine learning.This method doesn't require fluorescent staining of cells,realized non-contact,non-labeling analysis and diagnosis of cell.The measurement of data was obtained in a completely non-destructive natural state and will not damage cells.The main research contents are as follows.First,a three-dimensional optical cell model(OCM)with a variety of organelle structures(nucleus,cytoplasm,mitochondria,and lysosomes)was established based on confocal microscopy images of real prostate cells,which makes up for the traditional cell model with a single structure and is too ideal.This model effectively solves the problem of easily doped impurities when obtaining P-DI pairs of cells during the experiment,and greatly enriches the diversity of P-DI pairs,which provides convenience for the establishment of cell diffraction pattern libraries in the future.Second,developed a diffraction imaging simulation system with the same optical structure as the Polarized-Diffraction Imaging Flow Cytometry(pDIFC)developed by the group.Based on the theory of discrete dipole approximation(DDA)and optical tracking software(Zemax),simulated diffraction patterns of different normal and cancerous prostate cells were obtained.By comparing the simulated diffraction image with the measured diffraction image,it was found that they are roughly consistent,which indicated the reliability and correctness of the simulation system.Third,the size of the nucleus was changed in OCM,and the influence of the nucleus volume on the diffraction image was simulated.The results show that the change of nucleus volume has a significant effect on the diffraction pattern,which is highly consistent with the experimental results.Fourth,based on GLCM texture features and machine learning data mining,the normal and cancerous prostate cells are classified by unsupervised learning and supervised learning.It was found that the support vector machine(SVM)in supervised learning can effectively solve the problem that GLCM parameters are difficult to distinguish different cell types due to multidimensional characteristics,and the classification accuracy of SVM can reach about 95%,while the cells classification accuracy of unsupervised learning was low.