Research On Automated Cervical Cytological Smears Interpretation Method

Cervical cancer is one of the malignant tumors and a serious threat to women’shealth, which is one of three cancers of highest incidence. According to theGLOBOCAN2012report, the new cases of cervical cancer in the global scope in2012was up to nearly527,000, and approximately265,000women died of cervical cancer. InChina in2012, the number of new cases of cervical cancer was up to62,000, and29,000women died. However cervical cancer have longer lesions period, which makesit able to be found in the early mostly. Once it can be treated timely, the probability tocure cervical cancer would be high. Hence cervical cancer screening for women’s healthis very important.The inspection technique based on cervical cytologic smear is the most effectiveand feasible screening technology to detect cervical precancerous lesions and earlycervical cancer. The traditional screening methods based on the artificial interpretationhave some problems such as heavy workload, high cost and low reliability and accuracycausing by physician specialty technical and subjective mood, which makes the researchon automatic recognition technology based on image processing great significance forcervical cancer prevention and treatment.The thesis takes cervical cytologic smear interpretation automation as the researchobjective and the simulation of cytologist interpretation of cervical smear as researchthought of automatic interpretation. Combined with cervical cytopathologicalknowledge, we use image processing, Ontology modeling and semantic reasoningtechnology to study some key technologies of cervical cytology smears automaticinterpretation. The main contents of this thesis include: coarse segmentation of cervicalcytology smears, separation of overlapping cells and single cell segmentation; smearimage features extraction with human cognitive consistency on color, morphological&texture; smear ontology modeling of image feature, cytologic features&cervicallesions and semantic mapping; cervical smear semantic interpretation modeling. Thisthesis breaks through the traditional research method using the classifieras interpretation tools and puts forward an interpretation method of cervical smearbased on semantic reasoning, which would propose a new approach for cervicalcytology smear interpretation automation.The main contributions have: ①To propose a cell separation point detection method based on cell overlap areacurve curvature and an overlapping cell separation method based on elliptic curvefitting. In former method, firstly the concave region would be detected by analyzing thepositive and negative value of curve curvature, and secondly candidate overlap concaveregions would be selected according to curvature size as well as overlap concaveregions could be judged by width and interval distance of concave regions; and finallythe final contact points of overlapping cell contour (that is separation point) could befound by finding the curvature pole of overlapped concave region. In latter method,firstly the points on the contour curve between separation adjacent points would be usedas fitting data to get the ellipse fitting by using ellipse fitting method based on leastsquare method, and the fitting ellipses too big or too small would be filtered out to getfitting ellipse used for extracting separation line; and then the arcs between separationpoints on ellipse fitting would be extracted as separation line which would be used toseparate overlapping cells; and finally the relations between separation cell area and celloverlap area would be analyzed to determine whether the separation is effective or not.The method can keep the original shape of cells as well as reduce the probabilityof under-separation and after-separation.②To propose gradient vector flow active contour model based on Polarcoordinate system PGVF, which can greatly improve the speed of segmentation in thepremise of ensuring the segmentation accuracy. Firstly, PGVF transfer the cell imagesfrom Cartesian coordinates to polar coordinates to calculate the radiation gradient vectorflow; and then put forward "sand suppression algorithm optimized RGVF model" forthe optimization of edge map to eliminate the interference of the impurity within cells toedge image. Shown in the comparison experiment with RGVF active contour modulus,under the premise of ensuring the segmentation of accuracy, the proposed methodincreases the segmentation speed more than five times.③To propose a feature extraction method based on linear geometric heat flowevolution, whose extracted contour irregularity and smoothness is better with humansubjective cognitive consistency. After doing geometric heat flow evolution with cellcontour curve, this method could extract irregular and smoothness of contour bycomparing the curve after evolution and the original contour curve. The experimentresults of Herlev cervical cell image data set show that the extraction irregular featureswould be obvious correlation with cervical cell lesion. ④To propose chromatin granules characteristic extraction method based onmathematical morphology. Firstly, the limited adaptive histogram equalization methodis used to enhance contrast degree of nuclear images; secondly the cumulative sizedistribution would be obtained by doing open operation with different scalesof morphological basic structure elements, and the chromatin particle size distributionwould be got by doing derivation to the cumulative size distribution, and thecorresponding scale to the particle size distribution of the maximum value would beused as the chromatin granules feature descriptor. The experiment results ofHerlev cervical cell image data set show that the extraction cell chromatingranules characteristics would be obvious correlation with cell lesion.⑤To propose an automatic interpretation method by simulating cytology doctorsinterpret cervical smears, which is consistent with the principle of artificialinterpretation, easy to understand and make full use of existing knowledge andexperience of cytology cytologists. Firstly,the method builds image feature ontology,cytologic features ontology and cytologic smear cytologic features&interpretationresults ontology in correlation of interpretation; and secondly brings forward ontologysemantic mapping method based on semantic rule reasoning; and finally sets up cervicalcytology smears automatic interpretation system model based on ontology&semanticlogics, and elaborates the model principle as well as analyzes the relevant interpretationrules and the judgment result merging method. The contrast experimental results with Knearest neighbor (KNN) classifier&support vector machine (SVM) on the Herlev ofcervical cell image data set show that the interpretation method of cervical smear in thethesis has a high sensitivity and specificity for cervical cell lesion.