Microcalcification Clusters Detection Based On Subspace Learning And Support Vector Machine

Breast cancer is one of the most common malignant tumors which increases the ratio to mortality among women. Early detection and treatment are pivotal to reduce mortality. At present, mammography has been acknowledged as the most important and effective tool for detection. The microcalcification cluster (Mcc) is one of the main signs of breast cancer, whose detection occupies a particularly important status in the early diagnosis. However, there is only about 3% information in mammograms can be seen. Even experienced doctors are difficult to detect Mcc in time and the best time for treatment will be delayed. With the rapid development of computer technology, Computer Aided Detection (CAD) on mammograms has become a research hotspot in the early diagnosis of breast cancer.In order to effectively detect Mcc on mammograms and assist doctors in diagnosing breast cancer earlier, several detection approaches are studied in this paper. Firstly, a method is presented to extract region of interest (ROI) in mammograms automatically, by using independent component analysis (ICA) for feature extraction and training support vector machines (SVM) for pattern classification. The method is simple, effective, and intelligent. It provides a new research idea for ROI extraction. Secondly, to improve the detection performance and rate, a novel framework for Mcc detection based on subspace learning and SVM is developed. Subspace is used to extract discriminant features of Mcc in mammograms, and SVM is trained for classification. Finally, results are unstable as subspace is subjected to the impact of noise easily. To overcome the limitation of a single subspace, a method based on selective ensemble SVM of Mcc detection is proposed. Compared with traditional ensemble methods, the proposed selective ensemble method can not only effectively improve the generalization capability of ensemble classifier, but also reduce the size of classifier. In a word, several machine learning methods are proposed for detecting Mcc, experimental results show that the proposed detection methods not only can effectively detect ROI and Mcc in mammograms with good practicability and robustness, but also reduce the false positive rate, which provide some new ideas for the research of CAD system in the breast cancer.