An analysis of the contribution of scale in mammographic mass classification

Mammographic masses are often classified according to their shape as round, nodular or stellate. These classifications are useful in the recognition of benign vs. malignant masses. In this preliminary study, a set of 30 mammograms are analyzed. The 2D shape contour of each mass is mapped to a 1D radial distance measure. The discrete wavelet transform is applied to each measure in which a full decomposition is computed. The root-mean-square of the coefficients in each scale is then computed. These values are used as input features in a statistical classification system. The discriminating powers of these features are analyzed via linear discriminant analysis. The classification system utilizes a conventional Euclidean distance measure to determine class membership. The classification rates are 87%, 93% and 80% when using Haar, Daubechies-3 and Daubechies-5 wavelets, respectively