Mammographic mass recognition using feature reuse and channel attention mechanism

Abstract

Mass recognition is one of the important steps in early diagnosis for breast cancer. However, one of the major issues in mass recognition is low sensitivity. In this study, a novel convolutional neural network using feature reuse and channel attention mechanism is proposed and evaluated for mass recognition in mammograms. To improve representational power of network, our model takes advantage of two advanced techniques: one is feature reuse for exploiting the potential of the network through directing connections, and the other is the channel attention mechanism for modelling the interdependencies between channels by integrating the channel-wise information. The performance of our model is evaluated on DDSM and the values of the evaluation are 92.05%, 91.25%, 91.25%, 92.71%, 0.91 and 0.96 in terms of accuracy, precision, sensitivity, specificity, F1_score and AUC, respectively. The experimental results demonstrate the effectiveness of the proposed model to mass recognition task in mammograms.