Ultrasound Breast Cancer Image Classification Based On BI-RADS

Breast cancer is one of the common diseases of modern women.With the increase of the rhythm of life and the increase of stress in life,the incidence of breast cancer in adult women is on the rise.Ultrasound is one of the important detection methods for breast cancer.Because of its low price and low environmental requirements,ultrasound is widely used in China.Correct judgment of the severity of breast cancer can help doctors to make appropriate treatment plans and increase the cure rate.Ultrasound breast cancer BI-RADS grading standard is widely used in the international classification of ultrasound breast cancer,according to this standard,ultrasound breast cancer is divided into 0-6 a total of 7 levels,each level has different characteristics of breast cancer to assist Judgment,and each level corresponds to a different treatment plan.The research in this paper is to use ultrasound breast cancer image information for BI-RADS classification.In the past ultrasound breast cancer research,the general approach was to classify benign and malignant ultrasound breast cancer.The method used was traditional machine learning methods,relying on the characteristics of manual extraction.A purely binary classification is of little clinical significance because doctors do not take treatment based on a simple benign and malignant classification.Physicians need a more detailed classification of breast tumors.In the past,the use of manually extracted features for classification has a high requirement for experience and professional knowledge.Manual extraction features are obvious,easy to extract,and simple features.The features hidden behind the image that are not easily noticeable by the naked eye are Not extracted by hand.In order to solve these problems,this paper uses the deep residual network to learn the features hidden behind the image and performs a more detailed BI-RADS classification than benign and malignant classification.The main content consists of two parts:Firstly,this paper use the deep residual network to learn potential information in ultrasound breast cancer images.Instead of using ultrasound breast cancer images directly as input to the deep residual network,we instead use the ROI rectangular regions in ultrasound breast cancer images as depth residuals.The input of the network is studied so as to achieve the purpose of removing redundant information and accelerating the learning efficiency.And this paper have improved the deep residual network to change the traditional SOFTMAX layer in the residual network to SVM for BI-RADS classification of breast tumors.Secondly,this paper use computer systems to extract tumor edge information based on the ROI rectangle region of breast cancer images,use the breast tumor edge information to calculate priori BI-RADS features and image texture features,and combine these features with features learned from deep residual network.Further improve the accuracy of ultrasound breast cancer grading.Through experiments,found that the features learned from the deep residual network and the manually introduced features together with the SVM classifier can achieve the better classification effect.