seqAFF-ResNet:面向新冠肺炎的诊断模型

seqAFF-ResNet:diagnostic model of COVID-19

新冠肺炎的计算机辅助诊断是一种实现智能化影像诊断、临床诊断及临床分型的方法,在新冠肺炎的辅助诊断过程中,图像的病灶区域与组织边界对比不明显,导致模型不能较好地关注病灶区域,对有效特征的提取不够充分。针对上述问题,提出一个新冠肺炎辅助诊断模型seqAFF-ResNet(sequential attentional feature fusion-residual neural network)。设计串行注意力特征融合(sequential attentional feature fusion,seqAFF)模块,该模块串联条带注意力特征融合(strip attentional feature fusion,SAFF)模块和全局局部注意力特征融合(global local attentional feature fusion,GLAFF)模块,获取图像的纹理信息以及全局和局部信息,弥补卷积神经网络对于细节特征提取能力的不足,使得模型可以更好地关注于病灶区域;构造深浅层特征融合(deep and shallow feature fusion,DSFF)模块,使用深层特征的语义信息来影响浅层信息,同时将浅层的空间信息传入深层特征中,使深浅层特征进行有效融合,捕获丰富的上下文信息,实现跨层注意力特征增强,使网络能够更好地定位病变区域。与残差神经网络(residual neural network,ResNet)相比,seqAFF-ResNet准确率提升了3.42%,精确率提升了3.53%,F1分数提升了2.77%,AUC值提升了0.9%,实验结果表明,所提模型可以提高新冠肺炎的识别准确率,且与同类模型相比具有更好的性能。所提方法为新冠肺炎的辅助诊断提供了有效的识别方法,对新冠肺炎的计算机辅助诊断具有重要意义。

Computer aided diagnosis of COVID-19 is used to realize the intelligent image diagnosis,clinical diagnosis and clinical typing.In the process of auxiliary diagnosis of COVID-19,the obscure contrast between the focus area of the image and the tissue boundary,results in that the model can not well focus on the focus area as well as insufficient extraction of effective features.To solve these problems,a supplementary diagnosis model of COVID-19,seqAFF-ResNet was proposed herein.A serial attention fea-ture fusion module(seqAFF)was designed,which connected strip attentional feature fusion module(SAFF)and global local attentional feature fusion module(GLAFF)in series to obtain texture information as well as global and local information of the image to compensate the lack of detail feature extraction ability of convolutional neural network,so as to focus on the lesion area better.Deep and shallow feature fusion module(DSFF)was constructed using the semantic information of deep features to influ-ence the shallow information,meanwhile the spatial information of the shallow layer was passed into the deep features to fuse the deep and shallow features effectively.As a result,rich contextual information and achieving cross-layer attentional feature enhancement can be captured,enabling the network to better localize the lesion area.Compared with ResNet,the accuracy rate of seqAFFResNet increases by 3.42%,the accuracy rate increases by 3.53%,the F1 score increases by 2.77%,and the AUC value increases by 0.9%.The experimental results show that the model in this paper can significantly improve the recogni-tion accuracy of COVID-19,exhibiting a better performance compared with similar models.The method proposed in this paper provides an effective identification method for the auxiliary diagnosis of COVID-19,and is of great significance for the computer-aided diagnosis of COVID-19.