“ Model based classification of linear structures in digital mammograms ” ( “ Automatic detection and model based classification of anatomically different linear structures in digital mammograms

Computer-based analysis of digitised x-ray mammograms enables the automatic detection of subtle signs of cancer. These include stellate lesions and architectural distortions, both of which are characterised by patterns of linear structures. The automated detection of such lesions in digitised mammograms is a challenging task due mainly to the similarity between the lesions and other naturally occurring patterns of linear structures within the breast. breast. Consequently current algorithms produce a large proportion of false positive detections. This paper focuses on the detection and classification of anatomically different types of linear structures to enable accurate detection of abnormal line patterns. We demonstrate the automatic detection of lines via a non-linear multi-scale directional ridge operator and present a method of modelling the cross-sectional intensity profiles of the linear structures. The model, which is based on principle component analysis (PCA), was trained on structures automatically detected by the ridge operator and labelled by an expert breast-screening radiologist into different anatomical types. We present reconstructions of example profiles that demonstrate the suitability of PCA for profile shape description. The results of leave-one-structure-out classification experiments performed using the radiologist-labelled data are given and the ability of the model to classify seven anatomically different mammographic linear structures is demonstrated.