Content-based image retrieval techniques have shown great value in computer-aided diagnosis of mammographic masses. Many existing approaches adopt several features to better characterize mammographic regions. However, most of them fuse features through feature concatenation or result-level combination, which cannot fully exert the discriminative power of different features and also sacrifices the overall computational efficiency. To address these drawbacks, we propose to utilize coupled multi-index for index-level feature fusion. Specifically, complementary local features are extracted from the same locations of mammographic regions. Then, they are separately quantized using the 'bag of words' (BoW) approach. Finally, quantized features are inserted into a two-dimensional inverted index, with each feature corresponding to one dimension. Experiments are carried out on a large dataset constructed from the digital database for screening mammography (DDSM). Results demonstrate that our approach not only achieves better retrieval precision and diagnostic accuracy than individual features do, but also improves the overall efficiency and scalability.