Semantic Label Prediction of Mammography Based on CC and MLO Views

Abstract

Mammography image usually contains two views in different orientations-CC and MLO. In current computer-aided mammography label prediction systems, most models either assess information from only a single view which increases the false-positive rate or comprehensively evaluate information from four views of two breasts without distinguishing between the two breasts. In this paper, we propose a semantic label prediction model of mammography based on co-consideration of CC and MLO views. Firstly, a DCN is used to classify and enhance dense images. Secondly, a MLLN is designed to generate semantic labels of mammography by fusing the features of CC view and MLO view. The experiment shows that our model improves the prediction mAP of 7.3% compared to the single view model and 23.3% compared to the four-view model, respectively.