Combining Registration and Abnormality Detection in Mammography

摘要

Usually, image registration and abnormality detection (e.g. lesions) in mammography are solved separately, although the solutions of these problems are strongly dependent. In this paper, we introduce a Bayesian approach to simultaneously register images and detect abnormalities. The key idea is to assume that pixels can be divided into two classes: normal tissue and abnormalities. We define the registration constraints as a mixture of two distributions which describe statistically image gray-level variations for both pixel classes. These mixture distributions are weighted by a map giving probabilities of abnormalities to be present at each pixel position. Using the Maximum A Posteriori, we estimate the deformation and the abnormality map at the same time. We show some experiments which illustrate the performance of this method in comparison to some previous techniques.