Statistical modelling of lines and structures in mammograms

Abstract

Computer-aided mammographic prompting systems require the reliable detection of a variety of signs of cancer. The emphasis of the work described is the correct classification of linear structures in mammograms. Statistical modelling, based on principal component analysis (PCA), has been developed for describing the cross-sectional profiles of linear structures, the motivation being that the shapes of intensity profiles may be characteristic of the type of structure. PCA models have been applied to whole mammograms to obtain images in which spicules, linear structures associated with stellate lesions, are emphasised. The aim is to improve the performance of automatic stellate lesion detection by concentrating on those structures most likely to be associated with lesions.