A mammogram registration technique dealing with outliers

In this paper, we present a new method for registering images in presence of abnormalities. By abnormalities, we mean variations of image intensity which are due to pathologies and cannot be corrected by registration. Our approach consists of characterizing them as outliers. This characterization is obtained in a Bayesian framework, by defining registration constraints as mixtures of distributions which describe statistically image gray-level variations on both inlier and outlier pixels. Thanks to an outlier map weighting these mixture distributions, we can also take proper advantage of some prior knowledge about the lesion location. We use synthetic images and mammograms to illustrate the properties of the method and to compare it with some classical ones