基于改进CENet的新冠肺炎CT图像感染区域分割

Infection regions segmentation of COVID-19 CT images based on improved CENet

针对新冠病毒感染肺部CT图像感染区域分割精度低,小目标分割困难等问题,提出了一种改进CENet分割模型。首先,在编码阶段加入注意力机制挤压和激励(SE)模块引入全局上下文信息,增强特征提取阶段的感受野,加大目标相关特征通道的权重,从而提高小目标的分割能力;其次,引入特征聚合模块(FAM),采用双线性插值的方法,融合了不同层次的图像特征,得到更具有判别能力的表达,进一步提高网络的分割精度。基于COVID—19—CT-Scans数据集的实验结果表明:该算法分割结果与真实结果之间的重叠率Dice值为74.32%,平均交并比(MIoU)为80.34%,灵敏度(Sen)为84.25%,特异性(Spec)为99.14%。与现有的多种分割算法相比,该方法能更好地分割出新冠病毒肺部感染区域。

Aiming at the problems of low segmentation precision of infected areas of lung CT images infected by COVID-19 and difficulty in segmentation of small targets,an improved CENet segmentation model is proposed.Firstly,the attention mechanism SE module is added in the coding stage to introduce global context information,enhance the receptive field in the feature extraction stage,increase the weight of target-related feature channels,and improve the segmentation ability of small targets.Secondly,the feature aggregation module(FAM)is introduced,and the bilinear interpolation method is adopted to fuse the image features of different levels,so as to obtain more discriminant expression and further improve the segmentation precision of the network.The results based on the COVID-19-CT-Scans data set show that the overlap ratio between the segmentation results and the real results of the proposed algorithm,the Dice value is 74.32%,the mean intersection over union(MIoU)is 80.34%,the sensitivity(Sen)is 84.25%,and the specificity(Spec)is 99.14%.Compared with many existing segmentation algorithms,this method can segment lung infection areas of COVID-19 well.