Research Of Mass Detection And Classification For Breast Cancer In Mammography

Breast cancer is one of the most commom malignant tumors among women in the world.Its morbidity and mortality have rapidly increased in recent years.Early detection and diagnosis are the key to reducing breast cancer mortality.Mass is one of the most common direct signs of breast cancer in mammograms.In this context,this thesis focuses on the detection of masses and the classification of benign and malignant lesions.This thesis is divided into two main parts: the detection of masses and the recognition of benign and malignant lesions.The detection process is based on a multi-view analysis of bilateral mammogram images.Many mass detection methods only use one side of the breast information,which generates more false positives.In this thesis,suspicious regions are firstly detected on single view based on multiple concentric layers(MCL)detect method.Then we match corresponding regions in MLO and CC views with arc-based method.In addition to extracting geometric features and texture features,similar features are also extracted in the matched pair of suspicious regions,effectively utilizing the information of two views.At last,multiple twin support vector machines recursive feature elimination(MTWSVM-RFE)was used for feature selection.A dual support vector machine(TWSVM)classifier was trained to reduce false positives.In the part of mass classification,we introduce a method integrates joint L21-norm minimizing regularization with a nonparallel twin support vector machine,which is called TWSVML21.The L21-norm regularization selects features across positive and negative classes with joint sparsity.Due to the addition of regular terms,it will lead to difficulty in solving the objective function.Therefore,an iterative method is proposed to solve the this problem.The preliminary results on mass classification and three benchmark datasets show the feasibility and effectiveness of our TWSVML21.The test results on the DDSM dataset show that the multi-view analysis can reduce the false positives and the classification of benign and malignant lesions based on TWSVML21 also can achieve the desired effect.