Automatic registration of mammograms using texture-based anisotropic features

In this paper, an automated registration framework is proposed to identify the differences between corresponding mammographic images. The deformation between a pair of mammograms is approximated based on the matching of corresponding features on two images. First, a novel technique is employed to match the breast boundaries, aiming to maximize the mutual information between their curvature maps. Then, we apply Gabor filters onto the interior region of breast image, and extract texture-based anisotropic features. The registration process is accomplished through the recovery of the deformation field, in which both the positional and orientational attributes of the landmarks are registered correctly. The proposed technique is evaluated on three pairs of image pairs selected from MIAS digital mammogram database. The experimental results show that our method successfully registers corresponding mammograms with little human intervention, and becomes a valuable tool for effective detection of breast abnormalities.