Multi-view content-based mammogram retrieval using dynamic similarity and locality sensitive hashing

Abstract

Content-Based Mammogram Retrieval (CBMR) methods using Multi-View Information Fusion (MVIF) have triggered a growing interest in the last years given their ability to help radiologists make the right breast-cancer related decision. To further improve the retrieval performance, this paper introduces an efficient MVIF-CBMR method based on late fusion that combines retrieval result-level of Medio-Lateral Oblique (MLO) and Cranio-Caudal (CC) views. The proposed method adopts a coupled multi-index with a dynamic distance to evaluate the similarity between mammograms, which allows to fully exert the discriminative power of the complementary of MLO-CC features. Furthermore, the ROI dataset signature indexing step uses a hashing technique to optimize the computational time for retrieving relevant images. Thus, the proposed method takes two query ROIs corresponding to two different views (MLO and CC) as input and displays the most similar ROIs to each view using a dynamic similarity assessment. The retrieved ROIs can therefore be analyzed according to their clinical cases for the final decision-making relative to the query ROIs. The experiments realized on the challenging Digital Database for Screening Mammography (DDSM) dataset have proved the effectiveness and the efficiency of the proposed method.