Abstract:The comparison of bilateral mammogram images is important for finding masses especially in dense breasts. However, most existing mammogram mass detection algorithms only considered unilateral image. In this paper, we propose a deep model called contrasted bilateral network (CBN) to take bilateral information into consideration. In CBN, Mask R-CNN is used as a basic framework, upon which two major modules are developed to exploit the bilateral information: distortion insensitive comparison module and logic guided bilateral module. The former one is designed to be robust to nonrigid distortion of bilateral registration, while the latter one integrates the bilateral domain knowledge of radiologist. Experimental results on DDSM dataset demonstrate that the proposed algorithm achieves the state-of-the-art performance.

From Unilateral to Bilateral Learning: Detecting Mammogram Masses with Contrasted Bilateral Network.