Study On Classification Of Benign And Malignant Breast Tumors Based On Convolutional Neural Network

Breast cancer has become one of the diseases that have a significant impact on women's health.With the increase of life stress and environmental degradation,the risk of adult women suffering from breast cancer is also increasing.Mammography is an important means of detecting breast cancer.The doctor can initially determine whether the breast has lesions by reading the mammography image.But this early screening mainly relies on doctors prior knowledge and subjective judgment,which can easily lead to missed diagnosis and misdiagnosis.Therefore,the key to improve the prevention and treatment of breast cancer is to improve the speed and efficiency of doctor's film reading,realize double film reading,and reduce the phenomenon of missed diagnosis and misdiagnosis caused by doctor's personal experience.Computer-aided diagnosis system provides assistant advice for doctors in detecting and diagnosing lesions,which can improve the efficiency of imaging doctors,and then contribute to the early screening and prevention of breast cancer.Traditional machine learning aided diagnosis and deep learning aided diagnosis are two technical routes of computer aided diagnosis at present.Machine learning mainly relies on manual feature extraction of image data and model training using extracted features.Its process is time-consuming,laborious and unstable.The assistant diagnosis technology based on deep learning can automatically extract features,and there is no unstable feature caused by the difference of doctors' prior knowledge.The main research work includes the following:(1)In this paper,the data of mammography are segmented and de-centralized.In this study,we used threshold segmentation technology to segment the background and breast tissue of mammography image,which is too large and only a small part of the area is mammary tissue.We redefined the size to make the image close to the required size of the model,so as to reduce the complexity of the model.The data in this experiment is based on two open datasets,DDSM and INbreast.The data acquisition machines and parameters of these two datasets are different,which makes the contrast,format and size of the image inconsistent,leading to the generation of multi-center data.Multicenter data will seriously affect the accuracy of classification experiment results.In this paper,histogram specification is used to de-centralize multi-center data.After the pretreatment method proposed in this paper,the experimental results are improved by nearly 10%.Residual Net(ResNet)model is used for training,and the accuracy of the model results is as high as 81.6%.(2)Classification of benign and malignant breast tumors.In this paper,ResNet is selected as the model of this experiment.In order to optimize the model and enhance the experimental results,the ResNet model is modified.The modified model is named ResNet-B,ResNet-C and ResNet-D respectively.This paper also uses SVM classifier to replace the original soft Max classifier of the model,and uses the method of transfer learning to initialize the model parameters.Finally,the accuracy of classification based on the improved three models is 89.8%,90.2% and 91.4% respectively.