The Research On Classification Of Breast Tumor Based On Ultrasound Signs Scoring Features

Breast cancer as one of the common malignant tumor diseases among female population worldwide,has a high incidence and mortality.Due to its non-radiation,low cost,real-time and portable advantages,ult rasound imaging has become an important method during the early detection and diagnosis of breast cancer.In recent years,the rapid development of computer-aided diagnostic systems has played an important role in improving the objectivit y and accuracy of breast clinical diagnosis.However,most of the current diagnostic systems have a large difference with clinicans in the use of characteristics and its results are hardly understood and accepted by doctors,which has seriously limited its promotion and application in clinical diagnosis.Aimed at the practicalit y of the diagnosis system,this paper proposes two new novel diagnostic methods of mammography based on expert prior knowledge and data mining machine learning technology,which provides understandable and acceptable auxiliary decision for practical clinical diagnosis.First of all,there is a huge semantic gap between the traditional low-level features and the high-level semantic features understood by the doctors.In this paper,a quantitative scoring scheme for BI-RADS ultrasonic signs is proposed to obtain the tumor features that doctors can understand easily.Then,an improved biclustering algorithm was used to discover the expression patterns of benign and malignant tumors from a large number of medical data and transform them into effective diagnostic rules.Subsequently,Then,a rule-based ensemble learning method is used to establish the breast cancer identification model.Through the Ada Boost ensemble learning algorithem,the weak classifiers constructed by the benign and malignant rules are organically combined into the final strong classifier.The operation mechanism of proposed model is similar to the doctor's diagnosis method,so it has good interpretabilit y and generalization abilit y.In ad dition,aimed at the fact that the benign and malignant tumor has different misdiagnosis cost in clinical diagnosis,this paper presents a classification model based on a cost-sensitive BPNN method.The model fisrtly extract the abstract features for train ing neural network using biclustering based rules and then transform the network probability estimation into misclassification cost estimation based the minimum risk cost decision rule,thus the decision ouput with the least cost is obtained.In this paper,a large data set containing 1062 breast tumors was used to validate the proposed algorithms and the experimental result has been compared with the traditional methods.The experimental result shows that the proposed two methods have achieved good diagnostic performance,which provides a new idea for the auxiliary diagnosis system in the clinical field.