A Classification Algorithm for Breast Masses Based on the Combination of Bidirectional ConvLSTM Network and Coordinate Attention Module
Fangfang Chen, Xiaojin Wang, Yanna Zhang, Yanling Wang · 2023
To address the problem of breast mass size difference, in order to make the model pay better attention to breast mass information, this paper proposes a benign and malignant classification algorithm for breast masses based on the combination of bidirectional ConvLSTM network and coordinate attention mechanism, so as to meet the medical clinical needs. Firstly, the mass region obtained through multi-view mass detection is cropped as the input of the two-branch classification network, and the feature information of the mass region in different views on the same side is extracted separately; secondly, the SE sub-module of the MBConv module in the EfficientNet network is replaced by the coordinate attention module to enhance the expressive ability of the network to learn small masses; then, the bidirectional ConvLSTM module as the feature fusion module to connect the correlation between the two views, which then realizes the benign and malignant mass diagnosis. After experimental validation on the DDSM public dataset, it is concluded that the method proposed in this paper can obtain the values of accuracy and AUC of 0.910 and 0.963, respectively, which can effectively improve the benign and malignant mass classification accuracy.