Mass retrieval in mammogram based on hashing theory and linear neighborhood propagation

Abstract

Mass detection in mammograms usually has high false positive (FP) rate. Content based mass retrieval can effectively reduce the FP rate by comparing the image which is to be determined with mass images which have already been diagnosed. In this paper, a method combining discriminating anchor graph hashing (DAGH) and linear neighborhood propagation (LNP) is proposed for mammogram mass retrieval. Original AGH image representation does not consider pathological relevance in defining image similarity. To solve this problem, DAGH is put forward as a new image representation, which introduces the pathological class into image similarity. Furthermore, LNP is employed as a relevance feedback technique. Finally, interactive retrieval for mammogram masses is implemented based on the learning strategy between the underlying features and high-level semantic for images. Mammograms provided by the Breast Center of Peking University People's Hospital (BCPKUPH) are used to test the proposed method. Experimental results show that the DAGH image representation introducing pathological class is superior to original AGH in analyzing the similarity of mass images. Compared with existing methods, the proposed method shows obvious improvement in mass retrieval performance.