摘要
目的:将14种常用的机器学习分类模型应用于新型冠状病毒肺炎(COVID-19)与社区获得性肺炎(CAP)的鉴别诊断,探讨能为COVID-19的早期诊断提供最优效能的机器学习分类模型。方法:搜集经临床确诊的86例COVID-19患者和100例CAP患者的胸部CT图像,利用影像组学方法提取病变区域的纹理特征,使用14种机器学习分类模型构建放射组学特征,通过ROC曲线下面积(AUC)评估模型的诊断效能,AUC最高者的效能最优。结果:14种分类模型的AUC均大于0.9,随机森林(Random Forest)模型的AUC最高(0.9406),高斯贝叶斯(Gaussian NB)模型的AUC最低(0.9037)。结论:14种机器学习分类模型均可有效鉴别COVID-19与CAP,均具有较高的鉴别诊断效能(AUC值均>0.9),效能最高的模型是Random Forest(AUC=0.9406),能够在早期诊断COVID-19方面发挥优势。
Objective:Fourteen commonly used machine learning classification models were applied to the differential diagnosis of COVID-19 and community-acquired pneumonia(CAP)for exploring which can provide the best performance for the early diagnosis of COVID-19.Methods:Chest CT images of 86 clinically confirmed patients with COVID-19 and 100 patients with community-acquired pneumonia(CAP)were collected.The texture features of the lesions were extracted by radiomics methods.14 kinds of machine learning classification models were used to construct the radiomics features.The diagnostic efficiency of the model was evaluated by the area under the curve(AUC),and the diagnostic efficiency of the model with the highest AUC considered as the best method.Results:The AUC of the 14 classification models were all greater than 0.9,the AUC of Random Forest model was the highest(0.9406),and the AUC of Gaussian Bayesian model was the lowest(0.9037).Conclusion:All the 14 kinds of machine learning classification models can effectively differentiate COVID-19 from CAP,with high diagnostic efficiency(AUC>0.90).The most effective model is Random Forest,which can play out advantage in the early diagnosis of COVID-19.
作者
田斌
余晖
任基刚
汪汉林
徐井旭
黄陈翠
TIAN Bin;YU Hui;REN Ji-gang(College of Medical Imaging,Guizhou Medical University,Guiyang 550004,China)
出处
《放射学实践》
CSCD
北大核心
2021年第5期590-595,共6页
Radiologic Practice
