Automated benign and malignant breast masses classification is a crucial yet challenging topic. Recently, many studies based on convolutional neural network (CNN) are presented to address this task, but most of these CNN-based methods neglect the effective global contextual information. Moreover, their methods do not further analyze the reliability and interpretability of CNN models, which does not correspond to the clinical diagnosis. In this work, we firstly propose a novel multi-level global-guided branch-attention network (MGBN) for mass classification, which aims to fully leverage the multi-level global contextual information to refine the feature representation. Specifically, the MGBN includes a stem module and a branch module. The former extracts the local information through standard local convolutional operations of ResNet-50. The latter embeds the global contextual information and establishes the relationships of different feature levels via global pooling and Multi-layer Perceptron (MLP). The final prediction is computed by local information and global information together. Then, we discuss the reliability and interpretability of our mass classification network by visualizing the coarse localization map through Gradient-weighted Class Activation Mapping (Grad-CAM), which is important in clinical diagnosis. Finally, our proposed MGBN is greatly demonstrated on two public mammographic mass classification databases including the DDSM and INbreast databases, resulting in AUC of 0.8375 and 0.9311, respectively.

自动化的良性和恶性乳腺肿块分类是一个关键而又具有挑战性的话题。最近，提出了许多基于卷积神经网络（CNN）的研究来解决此任务，但是大多数基于CNN的方法都忽略了有效的全局上下文信息。而且，他们的方法没有进一步分析CNN模型的可靠性和可解释性，这与临床诊断不符。在这项工作中，我们首先提出一种用于大规模分类的新型多级全局引导分支注意网络（MGBN），其目的是充分利用多级全局上下文信息来细化特征表示。具体地，MGBN包括主干模块和分支模块。前者通过ResNet-50的标准局部卷积运算提取局部信息。后者嵌入全局上下文信息，并通过全局池和多层感知器（MLP）建立不同功能级别的关系。最终预测由本地信息和全局信息共同计算。然后，我们通过梯度加权类激活映射（Grad-CAM）可视化粗略的定位图，讨论了质量分类网络的可靠性和可解释性，这在临床诊断中很重要。最后，我们提出的MGBN在包括DDSM和INbreast数据库的两个公共乳房X线摄影质量分类数据库中得到了充分证明，其AUC分别为0.8375和0.9311。最终预测由本地信息和全局信息共同计算。然后，我们通过梯度加权类激活映射（Grad-CAM）可视化粗略的定位图，讨论了质量分类网络的可靠性和可解释性，这在临床诊断中很重要。最后，我们提出的MGBN在包括DDSM和INbreast数据库的两个公共乳房X线摄影质量分类数据库中得到了充分证明，其AUC分别为0.8375和0.9311。最终预测由本地信息和全局信息共同计算。然后，我们通过梯度加权类激活映射（Grad-CAM）可视化粗略的定位图，讨论了质量分类网络的可靠性和可解释性，这在临床诊断中很重要。最后，我们提出的MGBN在包括DDSM和INbreast数据库的两个公共乳房X线摄影质量分类数据库中得到了充分证明，其AUC分别为0.8375和0.9311。