基于通道注意力与空洞卷积的胸片肺气肿检测算法

Chest X-ray emphysema detection algorithm based on channel attention and dilated convolution

胸片中的肺气肿检测算法在临床辅助诊断中具有重要研究意义.针对已有算法缺乏特征通道筛选能力,特征图感受野较小易受局部组织噪声干扰,以及难易样本不均衡等问题,提出了一种基于通道注意力与空洞卷积的胸片肺气肿检测算法EDACD.首先,利用通道注意力模块构建了具有通道选择能力的特征提取网络SE-ResNet及特征金子塔网络SE-FPN.同时,使用空洞卷积代替了SE-ResNet中部分普通卷积,通过增大特征图感受野,提高所提特征鲁棒性.最后,采用焦点损失作为分类损失函数使网络专注困难样本训练.此外,训练过程中采用限制对比度自适应直方图均衡化算法对胸片进行预处理,进一步突出了肺气肿特征.通过数据扩充、聚类标签优化先验框参数的方式,克服肺气肿标注数据稀缺及传统先验框设置不符合肺气肿尺度等问题.在ChestX-Det10和ChestX-Det14数据集中的主、客观实验表明,所提算法较对比算法具有更好的肺气肿病灶检测能力.

Emphysema detection algorithm in chest X-ray plays an important role in clinical aided diagnosis.In order to solve the problems in existing algorithms,such as lacking feature channel screening ability,small receptive field of feature map,easy to be disturbed by local tissue noise,sample imbalance,EDACD(a chest X-ray emphysema detection algorithm based on channel attention and dilated convolution)is proposed in this paper.Firstly,the feature extraction network SE-ResNet and the feature pyramid network SE-FPN with channel selection ability are constructed by using channel attention module.At the same time,the dilated convolution is used to replace some common convolutions in SE-ResNet,and the robustness of the features is improved by increasing its receptive field.Finally,the focus loss is used as the classification loss function to make the network focus on training difficult samples.In addition,in the training process,the limited contrast adaptive histogram equalization algorithm is used to preprocess the image,which further highlights the characteristics of emphysema.Through data expansion and clustering labels to optimize the anchor parameters,so as to overcome the scarcity of labeled emphysema data and the inappropriate traditional anchors setting of emphysema.The subjective and objective experiment results in Chestx-Det10 and Chestx-Det14 datasets show that the proposed algorithm has better detection ability than the contrast algorithms.