高级检索

面向新型冠状病毒肺炎CT图像识别的DL-CTNet模型

DL-CTNet Model for CT images Recognition of Coronavirus Disease 2019

  • 摘要: 目前, 新型冠状病毒肺炎(COVID-19)仍处多地多点爆发形势. 利用深度学习技术辅助医务人员安全且高效地检测感染者是一种有效途径. 针对新冠感染者CT图像的磨玻璃影、铺路石征、血管扩张等特点, 设计了一种可有效提取CT图像中的局部与全局特征的轻量级网络:DL-CTNet. 输入预处理的CT图像后, 首先并采用空洞卷积和动态双路径多尺度特征融合(D-DMFF)模块两个支路提取浅层特征; 然后使用局部与全局特征拼接模块(LGFC)中的D-DMFF模块提取局部特征、Swin Transformer提取全局特征, 并通过拼接获得深层特征; 最后经过全连接层输出分类标签. LGFC模块以及DL-CTNet的低复杂度与有效性在两个CT图像数据集上得到验证. 实验结果表明, DL-CTNet的分类准确率高达98.613%, 与其他方法相比, DL-CTNet能更准确地识别COVID-19的CT图像.

     

    Abstract: Currently, coronavirus disease 2019(COVID-19) has not been effectively controlled. Deep learning methods can assist medical personnel in diagnosing COVID-19 safely and efficiently. According to the characteristics of ground-glass opacity, crazy paving sign and vasodilatation in CT images of COVID-19 patients, a light weight network DL-CTNet is designed to effectively extract the local and global features in CT images. Firstly, after inputting the pre-processed CT images, the shallow features are extracted by two branches of cavity convolution and dynamic dual-path multi-scale feature fusion(D-DMFF). Then, the D-DMFF module in local and global feature concatenation module (LGFC) is used to extract local features, and Swin Transformer is used to extract global features, and the deep features are obtained by the above two branches in LGFC. Finally, the classification label is output by the fully connected layers. The low complexity and effectiveness of CTNet are verified on two CT image datasets. The experimental results show that the classification accuracy of DL-CTNet is as high as 98.613%, and compared with other methods, DL-CTNet can more accurately detect COVID-19 CT images.

     

/

返回文章
返回