基于CCT的新冠肺炎胸部CT检测识别

Detection of COVID-19 Chest CT Images Based on CCT

根据国家卫生健康委员会公布的诊疗方案第五版,计算机断层扫描(CT)影像临床诊断结果可作为新冠肺炎(COVID-19)病例诊断的标准。CT图像能够清晰、立体地显示新冠肺炎患者肺部病变特征,针对新冠肺炎的诊断,可以使用胸部CT图像构建新冠肺炎检测模型,为医生提供更精确的诊断。本文提出了基于紧凑型卷积Transformer(CCT)的检测识别模型,首先使用U-Net分割网络提取肺区后,使用CCT对肺区进行识别。将Transformer编码器的注意力机制更改为了轴向注意力机制,并添加位置偏移项,在训练中获取更精确的上下文信息。在CC-CCII数据集中挑选出了1034张新冠肺炎CT图像,1003张社区肺炎CT图像和931张正常CT图像组成测试集,性能达到了98.5%的准确率,98.6%的灵敏度,并且在其他小型数据集上性能表现良好。证明了提出的方法使用胸部CT图像检测新冠肺炎有正向辅助作用。

According to the fifth edition of the diagnosis and treatment plan published by the National Health Commission of the People’s Republic of China, the clinical diagnosis results of computed tomography(CT) imaging can be used as the standard for diagnosis of COVID-19 cases. CT images can clearly and three-dimensionally display the characteristics of the lung lesions of patients with COVID-19. For the diagnosis of COVID-19, chest CT images can be used to construct a COVID-19 detection model to provide doctors with a more accurate diagnosis.This paper proposes a detection model based on the compact convolutional Transformer(CCT). First, the U-Net segmentation network is used to extract the lung area, and then the CCT is used to identify the lung area. The attention mechanism of the Transformer encoder is changed to the axial-attention mechanism, and the Positional bias term is added to obtain more accurate context information during training. In the CC-CCII data set, 1034 CT images of COVID-19 were selected, 1003 CT images of community pneumonia and 931 normal CT images were formed into the test set. The performance reached 98.5% accuracy and 98.6% sensitivity. The performance on the data set is good. It proves that the proposed method uses chest CT images to detect COVID-19 and has a positive auxiliary effect.