改进EfficientNet网络的COVID-19 X光分类

COVID-19 X-ray Classification Based on Improved Efficientnet Network

针对新冠感染蔓延速度快、新冠医学图像的人工诊断耗时长、压力大、医疗资源不平衡等诸多问题,在轻量级网络EfficientNet-B0的基础上引入新的注意力模块ECBAM,提出一种EfficientNet-ECBAM网络。首先用此模块替换EfficientNet-B0网络结构中SE模块能够改善其因降维操作导致部分细节丢失的问题。其次,因为ECBAM模块能在通道和空间2个维度进行特征提取,亦能改善SE模块提取图片特征信息不充分的问题。在选用的COVID-19胸部X光数据集上进行的实验得出,基于EfficientNet-B0网络改进后的EfficientNet-ECBAM网络的准确率比经典卷积神经分类网络VGG16、ResNet-50分别提升了3.76个百分点和2.13个百分点,特异性及敏感性等指标也均有提升,模型参数量也分别降低了97.3%、85.6%,比轻量级网络SqueezeNet、MobileNet V1的准确率分别提升了2.97个百分点和2.44个百分点。在消融实验中,改进的ECBAM模块的各项指标也优于其他注意力模块。实验结果表明,本文提出的EfficientNet-ECBAM网络模型具有分类性能好、参数量低、计算量小的优势,利于部署在经济欠发达地区的医疗机构。

In response to many problems such as the rapid spread of new coronary pneumonia,the time-consuming process of manual diagnosis using COVID-19 medical images,the imbalance of medical resources and the pressure of doctors’diagnosis,this paper introduces a new attention module ECBAM on the basis of the lightweight network EfficientNet-B0 and proposes the EfficientNet-ECBAM network.Firstly,replacing the SE module in the EfficientNet-B0 network structure with this module can improve the problem that some details of the SE module are lost due to the downscaling operation.Secondly,because the ECBAM module can extract features in both channel and space dimensions,it can also improve the problem that the SE module extracts insufficient information of image features.On the selected COVID-19 chest X-ray dataset,compared with the classical convolu⁃tional neural classification network VGG16 and ResNet-50,the accuracy of the improved EfficientNet-ECBAM network based on the EfficientNet-B0 network is improved by 3.76 percentage points and 2.13 percentage points respectively,specificity and sensitivity are also improved.The number of model parameters is also reduced by 97.3%and 85.6%respectively.Compared with the lightweight network SqueezeNet and MobileNet V1,the accuracy of EfficientNet-ECBAM is improved by 2.97 percentage points and 2.44 percentage points respectively.The improved ECBAM module also outperforms other attention modules in the ab⁃lation experiments in all metrics.The experimental results show that the EfficientNet-ECBAM network model proposed in this pa⁃per has the advantages of good classification performance,low number of parameters and low computation,which is favorable for deployment in medical institutions in less economically developed areas.