基于多通道双注意力网络的COVID-19图像分类

COVID-19 image classification based on multi-channeldual attention networks

针对逆转录聚合酶链反应对新冠肺炎(COVID-19)的检测存在一定的假阴性率、消耗时间过长等问题,提出了一种基于深度迁移学习的多通道双注意力网络(MDA-Net)对肺部图像进行检测。在深度迁移学习的框架下,引入了多通道双注意力模块,利用多个通道的位置关系,融合不同尺度的图像特征。将注意力机制和轻量级卷积神经网络相结合,扩大MDA-Net感受野,提高了对图像复杂区域和边缘区域的特征提取能力。MDA-Net在不同数据集上进行了实验,二分类任务和三分类任务分别能取得99.25%和99.39%的平均准确率,表现出良好的分类性能。

Aiming at the problems of a certain false negative rate and long time consumption in the detection of novel coronavirus pneumonia(COVID-19)by reverse transcription polymerase chain reaction,this paper proposes a multi-channel dual attention network(MDA-Net)based on deep transfer learning to detect lung images.Firstly,under the framework of deep transfer learning,a multi-channel dual attention module is introduced,which utilizes the positional relationship of multiple channels to fuse image features of different scales.Then,the attention mechanism is combined with a lightweight convolutional neural network to expand the MDA-Net receptive field and improve the feature extraction ability of complex and edge regions of the images.Finally,the MDA-Net is tested on different datasets,and the binary-classification task and three-classification task can achieve an average accuracy of 99.25%and 99.39%respectively,showing good classification performance.