218例老年患者孤立性肺结节的影像组学分析

Radiomics analysis of solitary pulmonary nodules in 218 elderly patients

目的探讨老年患者孤立性肺结节(SPN)的影像组学特征与病理的相关性,比较影像组学不同分类器模型及特征预测病理类型的准确性。方法回顾性分析2013年1月~2020年11月于解放军总医院经手术病理证实良恶性的218例SPN患者(良性组50例,恶性组168例)的术前薄层CT图像,标注病灶并进行图像预处理,提取病灶的影像组学特征,建立不同分类器的预测模型,评价模型的预测性能,与传统语义学特征预测结节病理类型的准确性进行比较。结果良、恶性结节的最大径、平均CT值的差异有统计学意义(P<0.05)。影像组学的形状特征参数平面度、伸长率、球形度及强度特征参数最大值、峰度、偏度、中位数、均方根的差异对鉴别结节的良恶性有意义。另外,有3个GLCM,5个GLSZM,1个GLDM,2个NGTDM纹理特征参数如群集阴影、区域百分比、相依熵、计算并返回对比度等在两组间的差别有意义。基于上述特征的Liner SVC分类器模型的预测效果最佳(AUC=0.8144)。恶性组中鳞癌组与腺癌组的最大径、平均CT值、空气支气管征的差异有统计学意义(P<0.05)。影像组学的形状特征参数最大2D直径(切片)及强度特征参数峰度、偏度、最大值的差异对鉴别鳞癌与腺癌有意义。有8个GLCM,2个GLSZM,2个GLRLM,4个GLDM纹理特征参数如相关性、尺寸区域非均匀标准化、短期重点、大依赖性高灰度级强调等在两组间的差别有意义。基于上述特征的Logistic Regression分类器模型的预测效果最佳(AUC=0.9521)。结论影像组学特征可以反映老年SPN不同病理类型间的差异,基于影像组学特征的分类器模型可以很好的鉴别不同病理类型的结节。

Objective To investigate the correlation between radiomics features and pathology of solitary pulmonary nodules in elderly patients,and to compare the accuracy of different classifier models and features in predicting pathological types.Methods The preoperative thin CT images of 218 patients with SPN confirmed by operation and pathology in our hospital were analyzed retrospectively,the lesions were marked and preprocessed,radiomics features of the lesions were extracted,the prediction models of different classifiers were established,the prediction performance of the models was evaluated,and the accuracy of predicting nodule pathological types with traditional semantic features was compared.Results There was significant difference in maximum diameter and average CT value between benign and malignant nodules(P<0.05).The differences of shape characteristic parameters such as flatness,elongation,sphericity and intensity characteristic parameters such as maximum,kurtosis,skewness,median and root mean square are significant in distinguishing benign and malignant nodules.In addition,there are 3GLCM,5GLSZM,1GLDM,2NGTDM texture feature parameters such as cluster shade,zone percentage,dependence entropy,calculation and contrast,which have significant differences between the two groups.The Liner SVC classifier model based on the above features has the best prediction effect(AUC=0.8144).There was significant difference in the maximum diameter,average CT value and air bronchial sign between squamous cell carcinoma and adenocarcinoma(P<0.05).The differences of shape characteristic parameters such as the maximum 2D diameter(section)and intensity characteristic parameters such as kurtosis,skewness and maximum are significant in distinguishing squamous cell carcinoma and adenocarcinoma.In addition,there are 8GLCM,2GLSZM,2GLRLM,4GLDM texture feature parameters such as correlation,size zone nonuniformity normalized,short run emphasis and large dependence high gray level emphasis,which have significant differences between the two groups.The Logistic Regression classifier model based on the above features has the best prediction effect(AUC=0.9521).Conclusion Radiomics features can reflect the differences among different pathological types of SPN in the elderly,and the classifier model based on radiomics features can well identify different pathological types of nodules.