A Study Of Novel Technique For Apoptosis Detection And Classification Based On Stain-free Diffraction Imaging Flow Cytometry

Apoptosis plays an important role in physiological or pathological processes.Conventional methods for analysis of cell apoptosis need special treatments for cells,which impact the integrity of the structure and function of the cell.In recent years,a stain-free method for apoptosis detection based on time-lapse phase-contrast microscopy images was developed,which can achieve an accuracy of around 90%.The other technique is photonic crystal biosensor imaging method for detecting the toxicity of drugs to cell populations.These techniques cannot be used for cell sorting after apoptosis induction.In this study,we developed a new method for stain-free cell apoptosis detection based on the polarization diffraction imaging flow cytometry and machine learning techniques,which can maintain the function and structural integrity of the cell and has potential for cell sorting as well.The main steps of our research work and results obtained are briefly described in the following.Cells were stained with Annexin V and DNA dye after apoptosis induction,and then were analyzed by flow cytometry and fluorescence activated cell sorting systems to obtain the three different subpopulations,which are viable cell,early apoptotic cell and late apoptotic/necrotic cell.The conditions were studied for diffraction image acqusition.Image analysis method was established to improve the quality of the image datasets.To obtain the best model for cell apoptosis recognition and classification,different texture extraction algorithms including gray level co-occurrence matrix,Gaussian Markov random field and uniform local binary patterns and machine learning method such as linear SVM,RBF SVM,logistic regression and random forest were studied and analyzed.Different models were built and compared.We obtained several models can achieve high precision and recall rates around 90% on test datasets.To further improve the model performance and reduce the time cost of the new method,the datasets generated by different algorithms were analyzed to show the features that have limited influence on model performance.Based on the results,the parameters in texture feature extraction algorithms were adjusted to generate new dataset with lower dimension.After the feature reselection,the model performance was estimated.The efficiency of the models before and after feature reselection was analyzed.The results show that the performances of the models were stable.The number of features in the GLCM model was reduced from 288 to 144,and the time for the model building was reduced from 8.66 ? 0.21 s to 4.54 ? 0.60 s.The number of features in the GMRF model was reduced from 146 to 78(K562 cells)and 74(HL60 cells),and the time for the model building was reduced from 13.01 ? 0.18 s to 7.02?0.17 s.The number of features in the LBP model was reduced from 2000 to 1280,and the time for the model building was reduced from 1.14 ? 0.03 to 1.12 ? 0.03.The time for the best classification model building based on these three different feature extraction algorithm were 17.12 ? 0.11 s,8.16 ? 1.05 and 36.85 ? 0.13 s.To further explore the feasibility of improving the classification accuracy,different features were combined for model building.But the results show that the combination of the features has no contribution to improving the accuracy of classification.This study successfully confirmed that the diffraction images of cells contain a lot of information,which can be used for non-staining classification of apoptosis after calculation and modeling.We obtained several models in this study.The model constructed by GMRF features and SVM has advantages in robustness,high efficiency and high accuracy.The model constructed by LBP features and SVM has the highest efficiency in all these models.It is more applicable to experiments in which efficiency is more important and desired than high accuracy.