Feature Extraction in Digital Mammography: An Independent Component Analysis Approach

In this paper we present a new feature extraction technique for digital mammograms. Our approach uses Independent Component Analysis to find the source regions that generate the observed regions of suspicion in mammograms. The linear transformation coefficients, which result from the source regions, are used as features that describe the observed regions in an effective way. A Principal Component Analysis preprocessing step is used to reduce dramatically the features vector dimensionality and improve the classification accuracy of a Radial Basis Function neural classifier. Extensive experiments in the MIAS database using a very small dimensioned feature vector gave a recognition accuracy of 88.23% when detecting all kinds of abnormalities which outperforms significantly the accuracy of the commonly used statistical texture descriptors.