Research On Classification Of Cervical Cells Based On Deep Neural Network

Cervical cancer,as one of the four major cancers with extremely high mortality in women,is a disease with definite cause and high early cure rate.Universal screening and treatment of early precancerous lesions of cervical cancer plays an important role in the prevention and treatment of cervical cancer.Cervical cytology,as the most important way for early screening of cervical cancer,requires manual classification and diagnosis by professional doctors.However,due to the lack of relevant professionals at present,it cannot meet the social demand for large-scale screening,and high-intensity work will affect the accuracy of manual classification and diagnosis.Therefore,it is of great value and significance to realize the research of automatic classification of cervical cell images by computer.The existing research schemes in the field of automatic classification of cervical cells are mainly divided into two categories: traditional algorithm and deep learning algorithm.The traditional algorithm process includes cell segmentation,artificial feature selection and cell classification.Because manual feature selection requires researchers to spend a lot of time adjusting parameters,most studies using traditional algorithms only divide cervical cells into normal cells and abnormal cells.Although it can provide some reference for doctors’ judgment,more information is needed for the diagnosis of the disease,such as the pathological types and pathological degrees of cervical cells.Deep learning has become the mainstream research method at present because of its ability to extract features automatically and it is not affected by the accuracy of cell segmentation and the perfection of artificial feature extraction.The main research content of this thesis is to use deep learning algorithm to realize multi-category accurate automatic classification of cervical cell images.The data sets used are Herlev data set and SIPa KMe D data set.The specific work contents are as follows:(1)Different algorithms have different emphases because of their different structures.In this thesis,depth-based multi-path convolutional neural network Dense Net121 and width-based multi-path convolutional neural network Xception are selected to conduct multi-classification experiments on Herlev data set and SIPa KMe D data set respectively.In the experiment,firstly,the data are amplified according to the characteristics of cervical cells.Then,according to the characteristics of each network,a single network is improved.For example,Drop Block module is introduced into Dense Net121 model to enhance the ability of anti-over-fitting,and SE(Squeeze-and-Exclusion)module is introduced into Xception model to enhance its ability of extracting cervical cell-specific features.The experimental results of multi-classification experiments on two data sets using Dense Net121 improved model and Xception improved model respectively show that the classification effect of the model is improved in different degrees compared with the network before improvement,and the total classification accuracy of multi-classification experiments on two data sets is above 98.5%.(2)In order to enhance the diversity of algorithm feature extraction and integrate the advantages of different algorithms,this thesis spliced the feature images output by the improved Dense Net121 network and the improved Xception network to construct a two-stream network DXNet.The experimental results show that compared with other fusion schemes,the fusion scheme of DXNet is the best fusion scheme,and the classification effect is better than that of the single network before fusion.Finally,DXNet achieves 99.84% and 99.88% accuracy on two data sets respectively,and all performance indexes have good performance,so it has a high recognition rate in the current cervical cells multi-classification algorithm.