基于多尺度特征选择网络的新冠肺炎CT图像分割方法  

COVID-19 CT image segmentation method based on Multi-scale feature selection network

作　　者：厉恩硕 张智诚 耿冰研 唐璐 Li Enshuo;Zhang Zhicheng;Geng Bingyan;Tang Lu(School of Medical Imaging,Xuzhou Medical University,Xuzhou 221004,China)

机构地区：[1]徐州医科大学医学影像学院,徐州221004

出　　处：《现代仪器与医疗》2023年第2期49-54,共6页Modern Instruments & Medical Treatment

基　　金：江苏省大学生创新创业训练计划(202110313037Z)。

摘　　要：目的改进U-Net图像分割算法,探讨多尺度特征提取对新冠肺炎CT图像分割应用的价值。方法利用Python编程语言和tensorflow和keras深度学习框架,在U-Net基线网络结构基础上,结合空洞空间卷积池化金字塔和高效通道注意力机制设计多尺度特征选择模块,提出改进后的多尺度特征选择网络(MsFS-Net),在COVID-19-20公开数据集上进行实验验证,使用5折交叉验证方法,对比U-Net、DeepLabV3+和Attention U-Net 3个网络模型,评测指标包括Dice、Recall、Precision,并绘制训练曲线图和可视化分割结果图。结果计算Dice、Recall、Precision 3项评测指标在5折交叉验证中的平均值和标准差,MsFS-Net的指标分别为(85.05±0.25)%,(85.33±0.20)%,(85.10±0.30)%,相比U-Net网络提升效果显著,平均值分别提升5.89%、5.71%、5.30%,标准差更小,训练曲线平滑稳定,对小病灶和不规则病灶的分割效果优于其他模型,模型鲁棒性更好。结论基于多尺度特征选择的图像分割方法具有高性能、高准确率的特点,在新冠肺炎图像自动分割中具有应用价值。Objective The value of multi-scale feature extraction for CT image segmentation of COVID-19 was discussed by improving U-Net image segmentation algorithm.Methods By using Python programming language,tensorflow and keras deep learning frameworks,Multi-scale Feature Selection module was designed on the basis of U-Net baseline network structure,combined with Atrous Spatial Pyramid Pooling and Efficient Channel Attention mechanism,and an improved Multi-scale Feature Selection Network(MsFS-Net)was proposed.The experiment was verified on the COVID-19 public data set,and 5-fold cross-validation method was used to compare the three network models U-Net,DeepLabV3+and Attention U-Net.The evaluation indexes included Dice,Recall,Precision,and so on.The training curve and visual segmentation result are drawn.Results The mean value and standard deviation of Dice,Recall and Precision in the 5-fold cross validation were calculated.The indexes of MsFS-Net were(85.05±0.25)%,(85.33±0.20)%and(85.10±0.30)%,respectively.Compared with U-Net network,the improvement effect was significant.The mean value increased by 5.89%,5.71%and 5.30%respectively,the standard deviation was smaller,the training curve was smooth and stable,the segmentation effect of small lesions and irregular lesions was better than other models,and the model had better robustness.Conclusion The image segmentation method based on Multi-scale Feature Selection has the characteristics of high performance and high accuracy,and has application value in the automatic segmentation of COVID-19 images.

关 键 词：新冠肺炎 医学图像处理 语义分割 多尺度特征提取 

分 类 号：R197.39[医药卫生—卫生事业管理] R445.3[医药卫生—公共卫生与预防医学]