Microcalcification Clusters Detection Based On Machine Learning

In digital mammograms, an important sign of the early breast cancer is the existence of microcalcification clusters (MCs). One of the key techniques for early diagnosis of the breast cancer is to detect MCs and to judge whether they are malignant or not in mammograms. However, there is only about 3% information in mammograms, which can be seen with the naked eye. Due to the most details in mammograms cannot been perceived by human eyes, it is even very difficult for an skillful radiologist to find the sign of early breast cancer, i.e., micalcification clusters, as a result missing the best time for treatment. To detect early sign of this disease and to aid doctors to diagnose breast cancer in early stage, we propose several new methods for enhancing and detecting the supicious areas by using some new techniques in machine learning, such as multi-resolution analysis, subspace learning, ensemble learning and so on. The main contributions of this paper are summarized as follows.(1) Because of the complex structure in mammography images, it is very difficult to choose typical training samples, which include the complex structure features, from mammograms. To overcome this difficulty, a new approach to MCs detection is presented based on active machine learning. In the proposed algorithm, firstly the microcalcification region is enhanced with a directional difference filter bank, which effectively extracts the features of MCs and suppresses the blood vessels and mammary duts. Then the active sample selection method based on Bootstrap is employed to select training set. Finally the trained Bayesian classifier can be used to detect MCs in mammogram. The experimental results show that the proposed algorithm reduces false positive rate with keeping the same sensitivity.(2) In order to improve the efficiency and the generalization ability of the MCs detector, a novel framework for MCs detection in mammograms is developed based on subspace learning and twin support vector machine (TWSVM). In the framework, MCs are firstly enhanced by using a simple-but-effective artifact removal filter and a well designed high-pass filter. Thereafter, subspace learning algorithms are embedded into this framework for subspace selection of each image block. Finally, the MCs detection procedure is performed in the feature subspaces. The experimental results show that MCs detection of the generalization ability and processing speed has been significantly increased. (3) For the purpose of making full use of the image spatial structure information, we generalized the vector-based learning algorithm, twin support vector machine (TWSVM) into the tensor-based method, twin support tensor machines (TWSTM) and successfully apply it to MCs detection. The experimental results show that the proposed algorithm achieved better detection performance than the TWSVM-based one, and could also deal the small sample problem better.(4) To get a better performance by ensembling the multiple methods for MCs detection rather than using a single algorithm, we developed a new ensemble learning method, Bracing, which has been applied to MCs detection in mammogram. In this method, when building new base learners the active relevance feedback is embedded to improve its generalization ability. Meanwhile, the weight of each base learner can be dynamically updated by the weighted score of the feedback result. Experimental results demonstrate that the Bracing algorithm could realize great advantages of ensemble classifier in generalization ability and could promise the preventing of overfitting.(5) As the subspace learning algorithm is sensitive to noise in training dataset, we proposed a hybrid subspace selective ensemble (HSSE), and successfully applied it to MCs detection in mammograms. In the algorithm, the subspace learning algorithm will be selectively ensembled according to the ability of preserving the classification information. Experimental results show that the proposed method improved the performance and stability of MCs detection and could be adapt to the noise enviroments better.In summary, to effectively detect the early sign of breast cancer in mammograms and to aid doctors to diagnose breast cancer better in early stage, we deeply studied the machine learning methods and their applications in microcalcification clusters detection. The proposed methods could get satisfactory results on sensitivity and reduce false positive rate, which provide some new ideas and methods for the research and development of computer-aided detection system in the breast cancer detection community.