Multi-classification Of Breast Cancer Histopathological Images Based On Convolutional Neural Network

Breast cancer has great harm to human health.In recent years,the incidence of breast cancer tends to be younger,and its morbidity and mortality are very high.Therefore,timely diagnosis and treatment can effectively control mortality.Histopathological examination is a key means of breast cancer diagnosis,so it is of great value to study the classification of histopathological images of breast cancer.Traditional classification methods require manual image preprocessing,feature selection and extraction of complex processes,not only timeconsuming,but also subjective diagnosis of different experts.In recent years,convolutional neural network has been widely used in medical image processing due to its strong feature learning ability.In this paper,based on Brea KHis database,convolutional neural network is used to achieve automatic multi-classification of breast cancer pathological images.The main research contents and innovations of this paper are as follows:(1)Current studies mainly focus on the judgment of benign and malignant breast cancer.However,different types of benign and malignant tumors require different treatment regimens,so it is of greater clinical value to further determine the specific types of breast cancer.In this paper,eight subtypes of breast cancer were classified based on benign and malignant classification,including adenosis(A),fibroadenoma(F),tubular adenoma(TA),phyllodes tumor(PT),ductal carcinoma(DC),lobular carcinoma(LC),mucinous carcinoma(MC)and papillary carcinoma(PC).(2)In order to overcome the traditional classification method is time-consuming,artificial difficult to extract features and tagged problems such as insufficient breast cancer data set,In this paper,two convolutional neural network models Goog Le Net and ResNet50 based on transfer learning are used as image feature extractors,and a parameter optimization strategy of global fine tuning is proposed to realize automatic classification of breast cancer,At the same time,data enhancement technology is used to further expand the data set to increase image diversity and reduce over-fitting.(3)Traditional transfer learning methods have unsatisfactory transfer effects due to large inter-domain differences.In addition,when sample categories increase,the similarity of cell morphological features between similar categories increases,resulting in feature learning difficulties and reducing accuracy of algorithm recognition.Breast cancer subtype eight categories of the identification accuracy is much less than benign and malignant,in order to solve this problem,in this paper,the method to improve the traditional transfer study,put forward a new model of transfer(in sample secondary transfer learning,ISSTL),using benign and malignant classification data sets as transitional data to optimize network parameters,and then using eight subtype classification data to train the network,further fine-tuning parameters,and experiments on Goog Le Net and ResNet50 models.It was found that ISSTL could significantly improve the classification accuracy of eight subtypes of breast cancer(Goog Le Net increased by 2.07%,ResNet50 increased by 2.57%).(4)Although the traditional convolutional neural network model has high accuracy,it also has many redundant features,which increases the computational burden of the model.In this paper,residual blocks of the model are improved according to the features of ResNet50 model to reduce the extraction of redundant features.In addition,the cross entropy loss function is improved by introducing weight adjustment factors to control the weight of samples that are difficult and easy to distinguish,so that the model focuses more on the samples that are difficult to classify during training.Experimental results show that the improved model(ResNet50-FL)has higher recognition accuracy and better generalization performance.The accuracy of binary classification is improved by 0.07%,and the model loss is reduced by 0.1211.The accuracy of eight classification is improved by 0.85% and the model loss is reduced by 0.3961.