Research On The Key Technology Of Computer Aided Diagnosis Of Mammography

The World Health Organization reports that breast cancer is one of the leading causes of death among women.The mortality rate of breast cancer can be reduced if breast cancer examination is carried out timely and accurately,and early and reasonable monitoring and treatment are carried out.Mammography is widely used as the first choice for early detection of breast cancer.The Computer Aided Diagnosis(CAD)system based on breast X-ray images can not only provide diagnostic reference for clinicians,reduce the workload of doctors in reading films,but also greatly improve the accuracy of early treatment of breast cancer.Statistical results show that about 80% of breast cancer patients are first diagnosed as a breast mass.Therefore,this paper takes breast tumor in mammogram image as the research object,discusses three key technologies of breast tumor CAD system: tumor region segmentation,feature extraction and benign and malignant tumor classification in order to construct a complete breast tumor CAD framework,and improve the accuracy and efficiency of the system.The main research results of this paper are as follows:(1)This paper proposed a network model based on multi-scale dilated convolution residual for breast tumor segmentation.Based on the classical structure of U-Net,the feature extraction module of multi-scale dilated convolution was designed which fused the dilated convolution with different expansion rates to enlarge the sensing range and obtain more local information on the premise of ensuring spatial information.The residual module was embedded,the batch normalization layer was added,and the binary cross entropy loss function multiplied by weight was introduced into the proposed network model.Therefore,the model retained the details of breast mass more completely without losing the image resolution,and the ability of extracting the regional features of breast mass was enhanced which finally improved the segmentation effect of the model.The segmentation model proposed in this paper was verified on CBIS-DDSM database,and the Dice coefficient and sensitivity obtained were 82.93% and 84.72%,respectively.Compared with U-Net model,it was increased by 0.75% and 1.36% respectively.(2)According to the characteristics of breast mass region,the spatial domain features and wavelet domain features were extracted.Fourteen spatial domain features were extracted including geometry,gray scale and texture.Six features of the mean and variance of wavelet coefficients at each scale were extracted after the three-scale wavelet decomposition.The above features can comprehensively and fully describe the relevant information of the tumor region,which was conducive to improving the accuracy of the classifier.(3)A SVM classifier optimized based on Genetic Algorithm(GA)was proposed to realize benign and malignant classification of breast masses.In view of the nonlinear separable situation of the classification task in this paper,RBF function was selected as the kernel function to achieve linear separable.GA algorithm was used to optimize the penalty parameter C in SVM and the parameter in RBF kernel function.The accuracy of the classifier was 91.80% and the area under the ROC curve was 0.9181 when verified on CBISDDSM database.Compared with typical SVM and other classification methods,the classification effect was better,indicating that the features extracted in this paper were effective and the designed classifier had good performance.