Benign Or Malignant Prediction Of Mammography Masses Based On Multimodal Feature Fusion

Breast cancer seriously threatens women's lives and health.Early detection and treatment will effectively improve the survival rate of patients.Mammography is the first choice for clinical screening of benign and malignant lesions.In the routine diagnosis of mammography,it is important to accurately identify the benign or malignant masses.Mass is a common symptom of breast cancer.The differences between benign and malignant masses in shape,edge and texture have become an important basis for radiologist to subjectively evaluate.However,the method relies heavily on the experience of the radiologist.Moreover,there exists great impact on the misdiagnosis of medical imaging lesions.Therefore,a tool for automatically diagnosing images is needed in clinical.It plays an important role in assisting radiologist to diagnose the mammography.With the Artificial Intelligence(AI)technology being actively employed in the medical field,breast tumor-assisted diagnosis has become a heated topic.Although AI technology has many advantages and has made great progress,the prediction algorithm for benign or malignant masses needs to be further improved in terms of multi-center,standardization and reliability.From the analysis based on hand-crafted features to the analysis of deep features,the study will improve the prediction of benign or malignant for breast masses using image biomarkers(hand-crafted features and deep features)in combination with clinical parameters.The main work of the paper is summarized as follows:Firstly,this paper predict the benign and malignant of breast masses based on traditional radiomics.The features of masses are complex and changeable.Radiomics technology is used to extract high-throughput features,aiming at increasing the separability of texture features between benign and malignant masses.This paper trains a classifier for discriminant analysis based on hand-crafted features.The SVM(The support vector machine)classification model effectively proves that the texture feature has the ability to predict the mass type.Next,considering that the results of traditional radiomics analysis need to be further improved,deep networks have advantages in the field of image diagnosis.A deep fusion network is proposed to predict benign or malignant masses of mammography.The study is based on the ImageNet pre-training network,a multi-level fusion structure is introduced toconstruct a VGG16 and Inception V3 fusion network,in order to explore the feasibility and accuracy of deep texture features for predicting benign or malignant masses.The results prove that the accuracy of the fusion model is higher than other basic networks.Finally,a method for predicting the benign or malignant mass of mammography images with multi-model feature fusion is proposed.The deep network is used as a feature extractor to extract deep features,merge hand-crafted features and clinical features,in order to increase the diversity of features.Moreover,drawing on the idea of stack generalization.The three types of features are respectively learned by the SVM model.The output results are fused by logistic regression.The study perform discriminant analysis.Finally,the prediction performance of the breast mass prediction algorithm is significantly improved.On the basis of the above theoretical method innovation,we design a benign or malignant masses prediction system for mammography.The system mainly implements image processing(Including image browsing,format type conversion,image enhancement,enlargement and reduction,measurement and other functions),breast cancer diagnosis,information extraction and storage.Moreover,it also includes a comparison of the prediction accuracy of the proposed algorithm.The system satisfies the need of clinical application.