Research On Assistant Diagnosis Of Breast Cancer Combining Cost-Sensitive And Semi-supervised Learning

Breast cancer is the second leading cause of cancer death in women, exceeded only by lung cancer. Digital mammography is one of the most important technologies in the early diagnosis of brease cancer. However, some lesions can not be found by radiologists due to the low quality images and people's neglect, so a computer aided medical diagnosis (CAD) system is needed to help diagnose. This paper presents some research on the pattern representation, feature selection and classification of CAD for the diagnosis of benign and malignant microcalcification clusters (MCCs).The first work of this paper is the extraction of 131 features to represente region of interest (ROI) based on the research of others. These obtained features can be divided into two kinds: 80 texture features and 51 spatial features (including Multi-scale features).Based on the first work, we modify the cost-sensitive selective ensemble (CSSE) algorithm to overcome its shortcomings such as low efficiency and instability. The modified CSSE (MCSSE) algorithm can select the most discriminate features by backward method. This algorithm reduces not only the relevant information and redundant information of feature subset, but also the computational complexity. Feature subset selected by MCCSE is more robust and discriminant than that selected by CCSE.In addition, semi-supervised learning is brought in the classification design of CAD. Traditional CAD system can not use the unlabeled samples got in general investigation of breast cancer. But semi-supervised learning can take sufficient advantage of the unlabeled samples. Considering the different cost of malignant and benign samples, we design classifiers with cost-sensitive learning and propose the semi-supervised cost-sensitive support vector machine.Besides above, Multi-view learning is introduced to classification design of CAD. There are two kinds of weakly correlated features in the digital mammography. Traditional CAD algorithms ignore multi-view and splice the features into a vector to classify. Differently, we bring a typical multi-view learning algorithm——Co-Training into CAD, and propose the CoCo-Training by improving the Co-Training algorithm according to the characteristic of cancer diagnosis. The experimental results show the superior performance of CoCo-Training algorithm.