Breast Cancer Tumor Diagnosis Based On Integrated Convolutional Neural Network

The incidence of breast cancer is increasing year by year,seriously endangering women’s health and life.The first symptom of most breast cancer patients is the presence of a mass in the breast,so breast mass is an early sign of breast cancer.Computer-aided diagnosis can provide objective reference opinions and reduce the burden on doctors.Therefore,it is of great value to study the classification of benign and malignant masses in mammography images combined with deep learning technology.At present,many scholars have made some achievements in the study of breast image classification based on deep learning,but the overall classification accuracy is not high.In this paper,a convolutional neural network model based on integration is proposed to improve the performance of breast classification.The main contents and innovations of this paper are as follows.(1)Constructed a breast image data set for model training.In view of the high resolution of mammography image and the small part of mass in the image,the mass area was standardized clipped in this paper.In order to solve the problems of blurred edges and low contrast of lumps in breast mammography images,median filtering was adopted in this paper to remove noise interference and retain edge information.In view of the insufficient number of samples of mammography target images,this paper carried out a series of data enhancement operations on the training data set,such as rotation,flipping,translation and scaling.The data set was expanded to 6 times the original size,which reduced overfitting to a certain extent.(2)Analyze and research three convolutional neural network models.Aiming at the complex problem of mammography target images,this paper designs an improved Dense Net model.The SE module in the SENet model is improved and embedded into it to form the SEDense Net model,and a variant of the SE-Dense Net model is formed according to the insertion position.Through experiments,the influence of embedding position on the classification performance of the model is studied,and the best embedding position is found.The pooling method of the model is also improved,and the GAMP pooling function is designed,which can not only extract the background information of the mammography target image,but also extract the texture information of the lump image.To solve the problem of large difference in the size of breast masses in mammography images,the Inception model was used to automatically select the convolution kernel of appropriate size,and the benign and malignant classification of breast masses was finally realized based on the Inception model.Experimental results show that the optimization model with Dropout operation reduces overfitting,and the influence of different optimizers on classification performance is also discussed.To solve the problem of a long training time and a large amount of computation,this paper chooses Shuffle Net model.By comparing with the classical model,it is verified that Shuffle Net model can reduce the computational complexity greatly while keeping the classification performance basically unchanged.This paper also explores the influence of the model grouping number on the classification performance,which provides relatively specific reference information for the construction of integrated convolutional neural network in the following chapters.(3)By combining deep learning with ensemble learning,an ensemble convolutional neural network model was designed to enhance the information exchange among classifiers and improve the classification accuracy of breast masses.The integrated neural network consists of three subnetworks with different structures fused by the integration strategy,which includes simple average strategy,weighted average strategy and majority voting strategy.Because of the different structure of the three subnetworks,they can extract the information of breast mass from different angles.After training and testing on CBIS-DDSM data set,the experimental results show that the integrated model has a better classification effect than the single model,and the classification performance of the weighted average strategy is better than the other two strategies by comparing the three different integration strategies.