改进全局上下文注意力新冠肺炎X光诊断方法

COVID-19 X-Ray Diagnosis Method Based on Improved Global Contextual Attention

在新型冠状病毒肺炎诊断任务中,胸部X射线(chest X-ray,CXR)的无关信息会影响模型分类决策,利用分割网络先将肺实质提取出来再进行分类是一种有效的途径。提出一种两阶段的分类模型Res-IgSa,使用ResUNet网络先提取CXR图像的肺实质,分类网络以ResNet50为基本框架,引入改进全局上下文模块(WGC)以及空间注意力模块(CSA)。WGC在全局上下文的基础上引入通道规范化以及门控自适应单元,以更好地调整通道间的关系;CSA引入分组卷积,通过分组数来控制模型的容量,WGC和CSA可以实现特征图间的互补。在COVID-19 Radiography DatabaseV5数据集上进行了大量的实验,与原论文采用相同的方法相比,Res-IgSa实现了更好的分类效果,准确率、精度、召回率以及F1值分别达到了94.154%、94.157%、94.154%以及94.139%,并进行消融实验以及结果可视化进一步证明了该模型的有效性。

In the task of diagnosing novel coronavirus pneumonia,the irrelevant information of chest X-ray(CXR)will affect the model classification decision.Segmentation network is used to extract lung parenchyma first and then classifica-tion is an effective way.A two-stage classification model Res-IgSa is proposed.The lung parenchyma of CXR images is extracted first by using the resunet network.The classification network takes ResNet50 as the basic framework,and intro-duces the improved global context module(WGC)and spatial attention module(CSA).WGC introduces channel normal-ization and gating adaptive unit based on the global context to better adjust the relationship between channels.CSA intro-duces group convolution,and controls the capacity of the model by the number of groups.WGC and CSA can realize the complementarity between characteristic graphs.A large number of experiments are carried out on COVID-19 radio data-baseV5 dataset.Compared with the same method used in the original paper,Res-IgSa achieves better classification effect,and the accuracy,precision,recall and F1 value reach 94.154%,94.157%,94.154%and 94.139%respectively.Ablation experiments and result visualization further prove the effectiveness of the model.