基于密度分布特征及机器学习诊断COVID-19相关性肺炎

目的:基于密度分布特征及机器学习诊断新型冠状病毒（COVID-19）相关性肺炎。方法:回顾性收集经荧光逆转录聚合酶链反应检测确诊COVID-19的患者42例（COVID-19组），社区获得性肺炎43例（对照组）。共获得211份胸部CT图像，以6:4比例分层抽样为训练集（126份）及验证集（85份）。采用一种CAD软件中的肺炎模块获得肺炎不同密度区间所占全肺体积的百分比（P/L%）。密度分布特征降维后采用支持向量机（SVM）建模，并评价4种核函数的SVM模型的诊断效能。结果:两组患者的年龄、性别及出现胸膜腔积液的构成比差异均无统计学意义（P>0.05）。肺炎密度分布特征降维后获得32个特征。基于该32个特征建立的4种核函数SVM模型中，多项式SVM模型在验证集的效能最高，受试者特征曲线（ROC）的曲线下面积为0.897（95%可信区间0.828~0.966），P<0.001。准确性为0.906（95%可信区间0.823~0.959），敏感性为0.906，特异性为0.906。结论:基于密度分布特征及机器学习诊断COVID-19相关性肺炎有较高的效能，有助于快速筛选COVID-19患者。

Diagnosis of COVID-19 associated pneumonia based on density distribution features and machine learning

Objective To diagnose corona virus disease 2019(COVID-19)associated pneumonia based on density distribution features and machine learning.Methods The clinical information of 42 patients with COVID-19 confirmed by RT-PCR(COVID-19 group)and 43 patients with community-acquired pneumonia(control group)were retrospectively collected.A total of 211 chest CT images were obtained,and according to stratified sampling based on a proportion of 6 to 4,the chest images were divided into training set(126)and validation set(85).The percentages of different density intervals of pneumonia in the total lung volume(P/L%)were obtained using a pneumonia module in CAD software.Support vector machine(SVM)was used for modeling after the dimensionality reduction of density distribution features,and the diagnostic efficiency of SVM models with 4 different kernel functions was evaluated.Results There was no significant difference in age,gender and constituent ratio of pleural effusion between two groups(P>0.05).A total of 32 features were obtained after the dimensionality reduction of pneumonia density distribution features.Among SVM models with 4 different kernel functions based on these 32 features,polynomial SVM model has the highest efficiency in validation set,and the area under receiver operating characteristic curve was 0.897(95%confidence interval 0.828-0.966)(P<0.001).The accuracy,sensitivity and specificity of polynomial SVM model were 0.906(95%confidence interval:0.823-0.959),0.906 and 0.906.Conclusion The diagnosis of COVID-19 associated pneumonia based on the density distribution features and machine learning has a high efficiency,which is helpful for the rapid screening of COVID-19 patients.