The final approach uses classification and has been inspired by the work of Karssemeijer and te Brake [84]. The d2 algorithm finds possible masses from the detection of spicules using second order Gaussian derivatives operators. If a line-like structure is present at a given site, the method provides an estimation of the orientation of these structures, whereas in other cases the image noise will generate a random orientation. With this information two new features are constructed. The first feature represents the total number of pixels pointing towards the centre, while the second one estimates if these directions are circularly oriented. With these two features and a set of classified mammograms d2 trains a binary decision tree. Subsequently, the decision tree can be used in segmenting unseen mammograms.