Detection Of Microcalcification Clusters Based On Multiresolution Analysis And Active Learning

Breast cancer is one of the most common malignant diseases, and the early diagnosis and treatment is the significant ingredient of any strategy designed to reduce breast cancer mortality. At present, mammography is the first choice to diagnose breast cancer. Clustered microcalcifications are one of the most important indications of malignancy in mammograms. Unfortunately, it is hard to recognize calcifications with naked eyes even in the high resolution mammograms due to the small size and various shapes of calcifications. With the rapid development of medical imaging modalities and computer technology, computer aided detection of microcalcifications in mammograms has become a hot topic for early diagnosis of breast cancer.In order to detect microcalcifications in mammograms efficiently, this paper presents an automatic detection system model which consists of three modules. Based on the discrete biorthogonal wavelet, a directional difference filter bank is introduced to enhance the mammograms, which can effectively suppress the linear component (LC) and enhance the microcalcification regions with nodular component (NC) simultaneously. Then the enhanced subimage is combined with the independent component analysis (ICA) to complete the description of the image, so the valid features for classification can be extracted. To effectively recognize microcalcification, the Bayesian classifier is trained by using the active learning sample selecting methods based on Bootstrap, which can guarantee the diversity of the negative samples and reduces the false positive rates simultaneously. The experimental results illustrate that the proposed detection modules with good practicability and robustness, can effectively detect regions of interest and microcalcifications in mammograms, which provides technical support for the computer aided detection of breast cancer.