Detection of Mammographic Masses by Content-Based Image Retrieval

Abstract

Computer-aided diagnosis (CAD) of mammographic masses is important yet challenging, since masses have large variation in shape and size and are often indistinguishable from surrounding tissue. As an alternative solution, content-based image retrieval (CBIR) techniques can facilitate the diagnosis by finding visually similar cases. However, they still need radiologists to identify suspicious regions in the query case. To overcome the drawbacks of both kinds of methods, we propose a CAD approach that integrates image retrieval with learning-based mass detection. Specifically, a query mammogram is first matched with a database of exemplar masses, getting a series of similarity maps. Then these maps are subtracted by discriminatively learned thresholds to eliminate noise. At last, individual similarity maps are aggregated, and local maxima in the final map are selected as masses. By utilizing a large database, our approach can effectively detect masses despite their variation. Moreover, it bypasses the identification of suspicious regions by radiologists. Experiments are conducted on 500 mammograms randomly selected from the digital database for screening mammography (DDSM) using receiver operating characteristic (ROC) analysis. The proposed approach achieves a promising ROC area index A _z  = 0.91, and outperforms two traditional classifier-based CAD methods.