Research On Automatic Diagnosis Methods Of Breast Cancer Based On Cost-Sensitive Learning And Its Application

Digital mammography analysis is one of the most important techniques for early detection of breast cancer. However, it is difficult to distinguish between benign and malignant microcalcification clusters (MCCs). A computer aided medical diagnosis (CAD) system could be helpful to the radiologists in assessment of MCCs. The CAD system usually contains three diagnosis steps: preprocessing stage, feature selection and classification. This paper follows such a three-stage model in theory and integrates the cost-sensitive learning into feature selection and classification respectively.We have done works below: first, the cost-sensitive selective ensemble (CSSE) algorithm is proposed and applied to select feature subset helpful for subsequent diagnosis. Since the calculated feature subset in the preprocessing stage may still contain redundant, correlated information which has the negative effect on the MCCs diagnosis usually, a feature selection is a crucial step of not only reducing the quantity of features and time cost but also, more importantly, boosting the diagnosis performance. Researchers have proposed numerous algorithms of feature selection; while seldom of them combine the feature selection with cost-sensitive learning to analyze MCCs'features. In fact, different features play different roles in MCCs diagnosis, which bring on various cost. Therefore, in this paper, we proposed Cost-Sensitive Selective Ensemble (CSSE) method to automatically select the most discriminative features of MCCs. The experiments show that CSSE is able to find a very helpful feature subset for classification of MCCs. Then, as the classification error costs are obviously unequal in practical MCCs diagnosis. On one hand, a false negative prediction may lead to fatal consequences, and on the other hand, a false positive prediction may be less serious, so we design a Cost-Sensitive automatic diagnosis system of breast cancer. This system uses CSSE to select features and Cost-Sensitive Support Vector Classifier (CSSVC) to classify. The experimental results on two mammogram databases (a benchmark DDSM and the ZhongDa databases (Developed by Jiangsu ZhongDa hospital)) show that CSSE performs better than Selective Ensemble (SE) without Cost-Sensitive on these problems. Moreover, Cost-Sensitive automatic diagnosis system of breast cancer represents better diagnosis performance than other systems in this paper.