Subjective and computer-based characterisation of mammographic patterns

We investigate the subjective characterisation of breast patterns and the possibility of automatic classification using computed measurements. Several clinical studies used mammogram pattern classification in order to investigate their relation to breast cancer. However, the establishment of an objective ground truth for pattern classification remains a problem that diminishes the potential of such studies since there is significant variability in the interpretation and coding of the patterns. For this reason, we investigate the observer variability in classifying mammographic patterns as well as the variability between film-screen and computer screen reading with respect to pattern classification. We also investigate the role of quantitative measures in characterising the breast parenchyma, discuss their agreement with the observer and argue that a normalised mammographic representation can improve the correspondence between measures and the perceived mammogram patterns.