A comparative study of image features for classification of breast microcalcifications

Computer-aided diagnosis systems for mammography have been developed in order to assist radiologists in the diagnostic process by providing a reliable and objective discrimination of benign and malignant mammographic findings. The effectiveness of such systems is based on the image features extracted from mammograms, which are mainly related to the morphology, texture and optical density of the suspicious abnormality. There are many methodologies reported in the literature able to provide a mathematical description of a mammographic lesion. In this paper, we apply various feature extraction methodologies on cases containing clusters of microcalcifications. Our purpose is to compare their performance in large scale in terms of classification accuracy and to investigate their potentiality in discriminating benign from malignant clusters. Experiments were performed on 1715 cases (882 benign and 833 malignant) extracted from the Digital Database of Screening Mammography, which is the largest publicly available database of mammograms. The results of our study indicated that texture features outperformed the rest of the considered categories, while the combination of the best features optimized the classification results, leading to an area under the receiver operating characteristic curve equal to 0.82.