Scalable mammogram retrieval using anchor graph hashing

Abstract

Mammogram analysis is known to provide early-stage diagnosis of breast cancer in reducing its morbidity and mortality. In this paper, we propose a scalable content-based image retrieval (CBIR) framework for digital mammograms. CBIR is of great significance for breast cancer diagnosis as it can provide doctors image-guided avenues to access relevant cases. Clinical decisions based on such cases offer a reliable and consistent supplement for doctors. In our framework, we employ an unsupervised algorithm, Anchor Graph Hashing (AGH), to compress the mammogram features into compact binary codes, and then perform searching in the Hamming space. In addition, we also propose to fuse different features in AGH to improve its search accuracy. Experiments on the Digital Database for Screening Mammography (DDSM) demonstrate that our system is capable of providing content-based accesses to proven diagnosis, and aiding doctors to make reliable clinical decisions. What's more, our system is applicable to large-scale mammogram database, such that high number analogical cases would be retrieved as clinical references.