Research On The Key Technologies Of Quantitative Analysis System For Cervical Cells

In recent years,the number of cervical cancer cases in developing countries is becoming younger and on the rise.At present,the effective means to prevent cervical cancer is early screening.The huge gap between the number of pathologists and the cases to be tested makes the large-scale implementation of cervical cancer screening in China meet a severe challenge.Many problems such as poor image focus resolution,inaccurate cell classification and segmentation,large quantitative calculation error,and low sample identification accuracy still exist in the quantitative analysis system of cervical cells.Aiming at the mentioned problems,this thesis proposes a Deoxyribonucleic acid(DNA)quantitative analysis system of cervical cells based on the existing work.This system obtains the images of cells by autofocusing and scanning the pap smears,then the cell segmentation and classification methods are carried on to acquire the lymphocytes and cervical epithelial cells.After that the epithelial cells are quantitatively calculated and corrected,finally,the time series data of cells are classified by Long-Short Term Memory networks(LSTM),to achieve the detection of cervical cancer.The main contributions of this thesis are as follows:(1)A microscope focusing system is designed and built.The focusing system is constructed by applying a microscope,3-d electric platform,and camera.The traditional focusing problem that the global image is clear while the epithelial cells are fuzzy has been greatly improved by the flow based on image preprocessing,extraction of cervical epithelial cells,definition calculation,and optimal focal plane searching.Experimental results show that the focusing flow proposed in this thesis can effectively enhance the image focusing effect.(2)The segmentation and classification of cervical cells are achieved by improving the existing model.A single cell image segmentation framework is proposed based on a watershed algorithm and improved Gradient Vector Flow(GVF)Snake model,which is optimized by modifying the gradient amplitude calculation method and adding the threshold processing step.By drawing from the existing work,the convex hull searching and curve fitting methods are applied to improve the segmentation boundary of overlapping cells,whose results are more consistent with the actual cell edges.The feature selection method based on the random forest is used to optimize the feature selection of morphology,chroma,texture,and optical density on the extracted cells.The classification accuracy reached 95.83% and 94.64% respectively in 2 and 5 categories,which are better than the other classification methods in the field.(3)The method of quantitative calculation for cervical epithelial cells is improved and LSTM model is used to classify the results of quantitative calculation to achieve the diagnosis target of cervical cancer.The cervical epithelial cell DNA values are corrected by the multiple nonlinear regression method,whose performance is demonstrated by the comparative analysis of mouse liver pap smears and references.The LSTM model is applied to classify the corrected time sequence data of cells to accomplish the accurate detection of cervical cancer.The accuracy,sensibility,and specificity of 98.3%,98.1%,and 97.9% are obtained respectively,which have certain advantages compared with similar methods.The experiment results prove that the cervical cell quantitative analysis system proposed in this thesis has certain advantages over the existing methods,and can partially solve the existing problems in cervical cancer detection,which has practical significance to be promoted.There are 48 figures,33 tables and 72 references.