Breast Cancer Classification Based On Deep Convolution Network

The incidence of breast cancer ranks first among female cancers and shows an increasing trend year by year.Therefore,the diagnosis and treatment of breast cancer have received more and more attention from the society.There are many types of tumor types of breast cancer,different characters of tumor correspond to different treatment plans,so the correct classification and recognition of breast cancer is particularly important in the process of breast cancer treatment.The traditional classification of breast tumor needs the help of imaging examination and pathological results to determine.The diagnosis process is cumbersome and inefficient.With the development of artificial intelligence technology,especially the deep learning and digital image processing technology,the computer diagnosis of breast cancer has become a research hotspot.This paper mainly studies the application value of magnetic resonance(MR)image combined with texture analysis method in the differential diagnosis of breast fibroma(FB),invasive ductal carcinoma(IDC)and invasive lobular carcinoma(ILC),as well as the application value of deep learning method in IDC pathological grade prediction.By analyzing the texture features of MR images and combining with convolutional neural network,we can predict different types of breast tumors,which can provide reference for breast cancer diagnosis.The main work and contributions of this thesis inclouds the following three aspects:(1)Based on texture analysis and multisequence MR image segmentation,a new method is proposed.In order to locate tumor area accurately,texture analysis method is applied to MR image segmentation.Using common texture analysis methods,such as wavelet transform and Gabor filter to obtain texture features of MR images,and according to the features,a feature classification model based on pixels is constructed.Then,the pixels are classified,and the common classification algorithms such as support vector machine and clustering algorithm are used to classify the image pixels,so as to achieve image segmentation.After getting the segmentation area,the edge of the tumor is extracted to complete the delineation of the focus area.At the same time,according to the characteristics of MRI image,clustering algorithm is used to cluster multiple sequences of MRI image.According to different clustering results,the segmentation of tumor region is completed and the edge of tumor region is extracted.Finally,compared with threshold segmentation,active contour model and other segmentation methods,the performance of the proposed method is verified.The experimental results show that this method can not only provide the basis for the feature extraction of different tumors,but also provide a reference for doctors to determine the target location in radiotherapy.(2)A model of breast tumor MR image classification based on Gabor feature is proposed.In order to judge the pathological types of breast tumors before operation,this paper makes full use of the advantages of computer technology in image recognition,and judges the types according to the differences between different tumor image textures.In this paper,firstly,Gabor wavelet is used to obtain the characteristic images in 5 scales and 8 directions.Secondly,the pixel mean value of the tumor area of the feature image is obtained by combining the sketch model,and the value is taken as the feature to construct the classification model.Then,according to the difference between different features,the key features are obtained.Finally,we use Bayesian,neural network,support vector machine and other classification algorithms to classify the above features,and compare the classification results.Accuracy,sensitivity and specificity were used to verify the performance of the model.The experimental results show that the accuracy of the model is between 69.81% and 77.36%.(3)Based on the deep convolution neural network and Gabor feature,a pathological grade prediction model of invasive ductal carcinoma was proposed.In order to predict the different pathological levels of invasive ductal carcinoma,Gabor wavelet is used to analyze MR image firstly.The texture features of MR image are obtained,and the classification model is built with support vector mechanism.The results show that Gabor texture features can distinguish two levels of invasive ductal carcinoma.Secondly,a convolution neural network model is constructed to classify two levels of invasive ductal carcinoma of the breast.By adjusting the parameters of the network structure,the accuracy of classification is continuously improved and the best parameter network is found.Then,the Gabor feature and convolution feature are combined to build a new feature layer,and SVM strategy is used to classify and complete the model construction.The experimental results show that compared with other methods,the classification effect of this method is significantly improved.The common three types of breast tumors can show their differences in MR images.The model proposed in this paper predicts the highest accuracy of the three types of breast tumors at 77.36%.At the same time,the highest accuracy in predicting two levels of IDC is 81.33%,indicating that breast MR images have important clinical value in identifying FB,IDC and ILC,and IDC levels.