Automated Feature Extraction from Breast Masses Using Multiscale Fractal Dimension

This paper proposes a new computer-aided diagnosis method to characterize benign and malignant masses in mammograms. First, for each image, a region of interest is segmented using an improved version of the EM/MPM algorithm. Then the contour is obtained by applying the Sobel high-pass filter. To extract the features, we compute the fractal dimension of the contour using the Bouligand-Minkoswki technique, with several successive dilations. This results in a curve, to which is applied multiscale differentiation, and twenty-two inflection points are obtained. The Y-axis coordinates are used to compose the feature vector. To select features, an attribute worth evaluator based on the 1R classifier is applied; it ranks three main inflection points that compose the final descriptor. This descriptor is submitted to the LADtree classifier, which finally suggests the diagnosis. The result of comparing the proposed method with traditional descriptors shows that it is well-suited to characterizing mammograms.