Researches On The Application Of Generative Adversarial Networks In Cervical Cell Image Classification

In the early screening of cervical cancer,the most commonly used method is cytology.The popularity of this method has greatly reduced incidence rate and mortality rate.Because of the heavy workload of the screening process and the shortage of professional doctors,computer-aided cytology can help doctors to judge the degree of cervical lesions.This technique is based on the automatic scanning of cervical smears.The evaluation criteria are objective and the results are of high clinical practicability and reliability.It is very helpful to improve the efficiency of cervical cancer screening.In recent years,deep learning has achieved good results in the field of medical image,and predecessors have done a lot of work in cervical cell classification.However,there is still room to improve the accuracy of classification,and the practicability of classification algorithm needs to be strengthened.In this paper,we combine the generation of network and classification network to improve the classification accuracy from many directions.The main work of this paper is as follows:1.In order to train convolutional neural network better when the data set is too small,a new sample is generated to enhance the image data by training GAN(Generative Adversarial Networks).The extended dataset is applied to the training image classification model,and good results are obtained.In order to solve the problem of two classification of Herlev cervical cell data set,this paper first uses the original training set to train Gan,generates a large number of high-quality high-resolution cell images,and expands each training set to 24000 cases.Then the expanded training set is used to train the classification network.The accuracy of the verification set in RESNET migration learning is up to 97%,which is higher than 93%of the training result of the data set expanded by affine transformation.It can be seen that this method can effectively realize the data enhancement of the image.2.In order to improve the accuracy of cervical cell image classification and improve the classification performance of the classification model under low resolution,this paper designs a joint classification network combining SRGAN super-resolution network and Res Net network using the cervical cell data set provided by Tongji Hospital etc.It realizes the work of super-resolution and classification at the same time,and updates the parameters of the whole network with the loss value of the final classification in the training process.The results of super-resolution image sampling and 5-fold cross validation show that the accuracy of multi classification of cervical cells in this network is greatly improved under the premise of high-performance super-resolution,which is 2.8% higher than the second method.3.In order to further improve the accuracy of cervical cell image classification,this paper combines convolutional neural network with manually extracted cell biological features,so that the final classification effect has been significantly improved.The absorbance of cell image is one of the important indexes to judge the degree of cell pathological changes.It takes the average absorbance of normal cells as the standard,so as to get the relative absorbance of other cells.In this paper,this feature is added to the full connection layer of Res Net network and the new network is trained.Compared with the original Res Net network,the classification accuracy is improved obviously.The results of 5 fold cross validation show that the total accuracy is 3.6%higher than that of convolutional neural network alone.