Research And Realization Of Cervical Cell Image Classification Recognition Method

Cervical cancer is one of the most common malignant tumors in current gynecological diseases,and it is also the only cancer that can identify the cause of all cancers at present.The cure rate of early cervical cancer can reach 100%.Therefore,early diagnosis and early treatment of precancerous lesions of cervical cancer are the main means to prevent and cure cervical cancer.With the popularization of cervical screening technology,the number of cervical smears is increasing day by day,resulting in the lack of cervical cytology screening staff and high pressure,or even due to fatigue and discomfort caused by factors such as the subjective interpretation of cervical smear errors.With the development of computer technology,the study of cytological screening of cervix by computer is becoming very important.Due to the complexity of cervical cytological lesions,professional screening physicians need to rely on years of medical clinical experience to accurately diagnose the disease.Therefore,it is a very difficult subject to make the final judgment of the disease with the direct help of computer technology,which is of little significance in the current stage.What is of research value is to provide effective auxiliary diagnostic information for professional screening physicians according to cervical cytopathological diagnostic criteria,such as morphological and texture characteristics of cells,prediction results of cell types,etc.Therefore,the main research contents of this thesis are as follows:1.Accurate segmentation of each cervical cell component.Aiming at the serious problem of improper segmentation of various components by FCN semantic segmentation network,this thesis designed a U-shaped semantic segmentation network called Herlev-Net for automatic segmentation of various components of cervical cells by combining the methods of unpooling and skip connection,and rough extraction of various components of cervical cells.In this thesis,the image texture information is integrated into the SLIC to obtain the fine edges of the components of cervical cells by the improved SLIC algorithm.Finally,the rough extraction results are fused with the fine edges.In this thesis,Pixel Accuracy and Mean Intersection over Union are used to verify that the proposed segmentation method has higher segmentation accuracy than existing algorithms.2.Extraction and importance analysis of cervical cell characteristic parameters.In this thesis,characteristic parameters of cervical cells were extracted from the three aspects of morphology,color and texture,and the importance of characteristic parameters was sorted by CART classification regression tree algorithm.8 most effective features were selected for the recognition of cervical single cell multi-classification.3.Cervical single cell multi-classification recognition.In this thesis,an ANNSoftmax classifier based on important features is designed firstly,which can realize a more accurate and efficient seven classification recognition of cervical cells.Then the structure and loss function of the semantic segmentation network Herlev-Net designed in this thesis are improved.A Herlev-CNet network was developed which could realize the segmentation of each cell component and the output of cell category.Through comparative experiments,it is verified that the improved Herlev-CNet network designed in this thesis can carry out more accurate seven classification recognition of cervical single cell image.4.Visualization of cervical cell characteristics and classification prediction result.In this thesis,the visual interface of characteristic parameters and classification prediction of cervical cells was designed and realized to provide auxiliary diagnostic reference for professional screening physicians.