Classification of linear structures in mammographic images

We have developed a novel technique for both the detection and anatomical classification of linear structures in mammographic images. Both stages are based on the statistical modelling of scale-orientation signatures. All experiments are based on a leave-one-out approach. The detection stage has been presented before [1]. For the classification stage real mammographic data was used. As the signatures incorporate scale aspects it is no longer essential to extract scale information which was previously shown to be a bottleneck in obtaining acceptable classification [2]. Spicule classification results show major improvements at low sensitivity and only moderate degradation at high sensitivity. We present spicule images based on both the detection and classification stages.