新型冠状病毒肺炎的影像组学研究

【摘要】为了识别新型冠状病毒肺炎（COVID-19）和非COVID-19肺炎（其他肺炎）的患者，提出一种基于胸部CT图像影像组学特征的分类方法。分别收集COVID-19患者和其他肺炎患者各90例的胸部CT图像，并手动勾勒肺炎病变区域；然后，利用影像组学方法提取病变区域的纹理特征和直方图特征，获得每个样本对应的一阶影像组学特征向量；最后，使用纹理特征和直方图特征作为输入，构建线性支持向量机（linear SVM）模型，对COVID-19患者和其他肺炎患者进行分类。该模型使用20次10折交叉验证进行训练和测试。对于COVID-19患者，还进行了相关分析（多次比较校正-Bonferroni校正，p<0.05/7），以确定纹理和直方图特征是否与血液的实验室测试指标相关。结果表明本研究提出的方法具有良好的分类性能，分类准确度高达87.56％，敏感度为82.78％，特异性为92.33％，受试者工作特性曲线下面积为0.939，这也证明了两组人群的影像组学特征是高度可区分的，此模型可以有效地识别和诊断COVID-19患者和其他肺炎患者。相关分析结果显示某些纹理特征与白细胞、中性粒细胞和C反应蛋白正相关，而也有某些纹理特征与血氧和中性粒细胞负相关。

Radiomics analysis on COVID-19

A classification method based on radiomics features of chest CT images is proposed to identify patients with COVID-19 and those with other pneumonias. The chest CT images of 90 patients with COVID-19 and 90 patients with other pneumonias are collected in the study, and the regions of interest of pneumonia are manually outlined. Then radiomics is used to extract the texture features and histogram features of the lesion regions, thereby obtaining the first-order radiomics feature vector of each sample. Finally, the texture features and histogram features are taken as inputs to construct a linear support vector machine model for classifying patients with COVID-19 and patients with other pneumonias. Ten-fold cross-validation is conducted 20 times for training and testing. For patients with COVID-19, correlation analysis(multiple comparison correction-Bonferroni correction,p<0.05/7) is also carried out to determine whether the textural features and histogram features are correlated with laboratory indexes of blood. The results show that the proposed method has excellent classification performances, with a classification accuracy up to 87.56%, a sensitivity of 82.78%, a specificity of 92.33% and an area under receiver operating characteristic curve of 0.939, which proves that the radiomics features of the two groups are highly distinguishable and that the proposed model can effectively identify and diagnose patients with COVID-19 and patients with other pneumonias. The correlation analysis results reveal that some texture features are positively correlated with white blood cell, neutrophils and C-reactive protein, and that there are some other texture features negatively relative with blood oxygen and neutrophils.