Mammogram segmentation using multi-atlas deformable registration.

Accurate breast region segmentation is an important step in various automated algorithms involving detection of lesions like masses and microcalcifications, and efficient telemammography. While traditional segmentation algorithms underperform due to variations in image quality and shape of the breast region, newer methods from machine learning cannot be readily applied as they need a large training dataset with segmented images. In this paper, we propose to overcome these limitations by combining clustering with deformable image registration. Using clustering, we first identify a set of atlas images that best capture the variation in mammograms. This is done using a clustering algorithm where the number of clusters is determined using model selection on a low-dimensional projection of the images. Then, we use these atlas images to transfer the segmentation to similar images using deformable image registration algorithm. Our technique also overcomes the limitation of very few landmarks for registration in breast images. We evaluated our method on the mini-MIAS and DDSM datasets against three existing state-of-the-art algorithms using two performance metrics, Jaccard Index and Hausdorff Distance. We demonstrate that the proposed approach is indeed capable of identifying different types of mammograms in the dataset and segmenting them accurately.