Improved Mammographic Mass Retrieval Performance Using Multi-View Information

Breast cancer is the most common malignant disease in women. Mammographic mass retrieval system can help radiologists to improve the diagnostic accuracy by retrieving biopsy-proven masses which are similar with the diagnostic ones. However, although screening mammograms usually consists of two-view(MLO and CC) mammography of the same breast, most breast CAD systems incorporate with image retrieval techniques are based on a single-view principle where query ROI within a view is analyzed independently. In this paper, a mammographic mass retrieval approach based on multi-view information is proposed. In this work, the query example is a multi-view(MLO and CC) mass pair instead of the single view mass in the traditional image retrieval framework. In the experiments, several visual features are used for retrieval evaluation. Both distance similarity measures, such as Euclidean distance, and k-NN regression model based nondistance similarity measures are used for comparison. Experimental study was carried out on a database with 126 biopsy-proven masses(63 mass pairs). Preliminary results showed that multi-view based retrieval approach achieves better retrieval accuracy than single-view based one, especially for the k-NN regression model based similairy metric.