Benign And Malignant Classification Of Nodules In Mammography Images

According to the survey results of the World Health Organization,the incidence of breast cancer ranks first in women’s malignant tumors,which seriously affects women’s physical and mental health.The symptoms and signs of breast cancer are diverse,and the symptom between women in Asia and Europe is different,which makes the diagnosis of doctors more difficult.In order to reduce the mortality rate of breast cancer in China,the research and design of a classification model of benign and malignant breast masses have very important theoretical and applied research values.Based on deep learning research,a benign and malignant classification method for breast masses was designed.Aiming at overcoming the problem of unclear boundary of various tissues in breast molybdenum target image,the connection between breast masses and other tissues,a non-linear data preprocessing method-histogram equalization,combined with truncated normalization method is used to generate a new image;aiming at overcoming the problem of huge difference in nodule target size,and in order to make the model pay more attention to breast mass information,on the basis of the output of the benign and malignant classification results of the mass,the segmentation results of the breast mass are added,and the network structure of Res Net50 is improved.For solving the problem of overfitting network model caused by the small amount of labeled data,label smoothing is used to suppress the confidence of the model in the prediction results and enhance the generalization ability of the model.In view of the inability to migrate from a natural image to a deep model of mammography,a two-stage learning and network layered learning methods are used to fully utilize the information of the existing data set and improve the accuracy of the benign and malignant classification model;for the problem that the training network is not easy to converge,the training methods of learning rate warm-up and learning rate decay are adopted.Tested on the DDSM data set,the benign and malignant classification of the tumor reaches an accuracy rate of 0.9887,and the area under the receiver operating characteristic curve(ROC)AUC reaches 0.9979.In addition,the accuracy on the INbreast dataset is 0.9565,and the AUC is 0.9867.The above experimental results show that the benign and malignant classification of the breast mass classification network model has a high accuracy,and a high reference value for the clinical research of breast cancer diagnosis.