MSAG-TransNet:肺部CT图像中新型冠状病毒感染区域的分割模型

MSAG-TransNet:Segmentation model of COVID-19 infected areas in lung CT images

针对现有新型冠状病毒感染区域的分割方法存在形态特征提取不充分、感染区域检测不完整以及背景混淆等问题,提出了一种肺部CT图像中新型冠状病毒感染区域的分割新模型:MSAG-TransNet模型。该模型在U型网络的基础上增加了多尺度特征抽取模块、Transformer语义增强模块和多重注意力门模块等3个新模块。首先设计了多尺度特征抽取模块来增强骨干网络的特征提取能力,通过多分支结构的深度可分离卷积,充分提取感染区域的形态特征;其次,设计了Transformer语义增强模块来捕获图像全局位置信息,整合局部形态特征;最后,设计了多重注意力门模块,将提取的特征与对应上采样过程的门信号拆分成不同分区,然后利用注意力门抑制各分区的无效特征,得到最终分割结果。该模型在两个公开的新型冠状病毒感染CT数据集上进行实验,实验结果显示:分割图像的Dice系数分别为82.03%和76.67%,精确率为77.27%和72.34%,交并比为69.53%和62.16%;与其他主流模型相比,该模型能够提取更丰富的形态特征,检测到更完整的感染区域,并且得到更精准的分割结果。该模型可以更精确的定位和量化新型冠状病毒感染区域,为临床诊疗提供可靠参考。

To solve the problems of insufficient morphological feature extraction,incomplete detection of infected areas and background confusion in the existing segmentation methods for infected areas of COVID-19,a new segmentation model MSAG-TransNet for infected areas of COVID-19 in lung CT images was proposed.This model adds three new modules to the U-shaped network:a multi-scale feature extraction module,a Transformer semantic enhancement module and a multi-attention gate module.Firstly,the multi-scale feature extraction module was designed to enhance the feature extraction ability of the backbone network,and a deep separable convolution of a multi-branch structure was used to fully extract the morphological features of the infected areas.Secondly,the global position information of the image was captured through the designed Transformer semantic enhancement module,and local morphological features were integrated.Finally,the multi-attention gate module was designed to split the extracted features and the gate signal corresponding to the upsampling process into different partitions,and then attention gates were used to suppress invalid features in each partition to obtain the final segmentation result.The model was tested on two publicly available COVID-19 CT datasets.Experimental results show that the Dice indexes of the segmented image are 82.03%and 76.67%,the accuracy rates are 77.27%and 72.34%,and the intersection and merger ratios are 69.53%and 62.16%,respectively.Compared with other mainstream models,this model can extract richer morphological features,detect more complete infected areas,and obtain more accurate segmentation results.Therefore,this model can more accurately locate and quantify the infected areas of COVID-19,and provide reliable guidance for clinical diagnosis and treatment.