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. Existing segmentation algorithms underperform due to variations in image quality and shape of the breast region. In this paper, we propose to segment breast region by combining data-driven clustering with deformable image registration. In the first phase of the approach, we identify atlas images from a dataset of mammograms using data-driven clustering. Then, we segment these atlas images and use in the next phase of the algorithm. The second phase is atlas-based registration. For a candidate image, we find the most similar atlas image from the set of atlases identified in phase one. We deform the selected atlas image to match the given test image using the Demon’s registration algorithm. Then, the segmentation mask of the deformed atlas is transferred to the mammogram in consideration. Finally, we refine the segmentation mask with some morphological operations in order to obtain accurate breast region boundary. We evaluated the performance of our method using ground-truth segmentation masks verified by an expert radiologist. We compared the proposed method with three existing state-of-the-art algorithms for breast region segmentation and the proposed approach outperformed all three in most of the cases.