Investigating the image features landscape for the classification of breast microcalcifications

Computer aided diagnosis systems using machine learning techniques have been developed in order to assist radiologists' diagnosis and overcome inherent limitations of conventional mammography. Such systems base their diagnosis on image features extracted from mammograms, which are mainly related to the shape, the morphology, the texture and the position of the suspicious abnormality. Since the discrimination of malignant and benign lesions is a classification problem, a feature selection preprocessing step is needed in order to minimize the dimensionality of the features set by keeping the most significant between them. In this paper, we compare four feature selection methods all based on different approaches on ranking and selection and perform classification of data. Experiments were performed on cases containing clusters of microcalcifications, extracted from a large public mammography database. Our findings indicate that there are subsets of very small number of features that can provide a proper baseline classification.