Study On Classification Of Breast Tumors Based On Convolutional Neural Network

The incidence of breast cancer has ranked the first place among all malignant cancers in women,and breast cancer has become the first killers of women’s health.At present,mammograms(MGs)is the main method for screening breast tumors.The clinical and imaging manifestations of benign and malignant breast tumors overlap partially,making it difficult to distinguish them.Therefore,it is urgent to establish an effective,intelligent and high-precision classification method for breast tumors.With the continuous development of smart healthcare,computer-aided diagnosis systems have emerged to help doctors more accurately identify the benign and malignant lesions.Most notably,convolutional neural network(CNN)has conducted in-depth research on the classification of benign and malignant breast tumors deeply,but the accuracy needs to be improved.Moreover,the latest Efficient Net-B0 model has not been studied.Therefore,CNN-based classification methods of breast tumor benign and malignant were proposed for in-depth research.The main contents included standardizing and preprocessing multi-center data,comparing the performance of the four CNN models on the MGs image classification,and fine-tuning and improving the optimal performance model.The details were as follows:First,due to the lack of MGs image data and imbalance of samples,the classification results are prone to overfitting.Therefore,this study proposed to use multi-center data to increase the sample size.The three open databases included MIAS database,DDSM database,and INbreast database.The images of the three databases were standardized and preprocessed,mainly including format and pixel unification,feature center cropping,histogram equalization,and data enhancement,etc.,which greatly increased the amount of data and improved the image quality for subsequent experiments.To illustrate the role of the extended data,this study analyzed the necessity and effectiveness of extended database using the Efficient Net-B0 model.Besides,model could typically avoid overfitting their data with sufficiently large training datasets.Then,the performance of VGG-16,Res Net-50,Inception V3 and Efficient Net-B0 models on MGs image classification were evaluated.In addition,each model used the Image Net data for transfer learning,and then were trained and tested by the extended database.The results showed that transfer learning could effectively shorten the training time,and the performance of each model has been improved.Importantly,the performance of the Efficientnet-B0 model on the classification of benign and malignant breast tumors was significantly better than other models.Finally,the Efficient Net-B0 model was further refined by fine-tuning learning rate,batch size,dropout and composite adjustment,and improved by adapting network depth,network width,image resolution,composite zooming using adaptive zoom technology and replacing SE module with Dense-shortcut module.The results showed that composite adjustment and zooming worked best.Additionally,the learning rate and“composite zooming + Dense-shortcut” were most sensitive to performance improvements.In summary,Efficient Net-B0 model was effective for the classification of breast tumors.The research results also have promotion value and practical implications for current artificial intelligence methods used for the classification of breast tumors.