摘要
目的基于胸部CT增强图像,进行纹理分析技术预测肺腺癌患者淋巴结转移的可行性研究。方法回顾性分析60例行术前常规胸部CT增强检查,并于2周内进行手术切除及系统性淋巴结清扫术,经病理证实的肺腺癌。根据术后病理将其分为两组,无淋巴结转移者25例及有淋巴结转移者35例。选取癌灶最大单幅层面,基于Mazda软件勾画感兴趣区。利用Fisher系数提取、交互信息提取(MI)、分类错误联合平均相关系数提取(POE+ACC)、三者联合提取(FPM)及1-NN最近邻算法,联合B11软件包提供的原始数据分析(RDA)、主要成分分析(PCA)、线性分类分析(LDA)及非线性分类分析(NDA)4种机器学习算法进行分析,结果以正确率显示。选取在5种提取方法中出现3次或3次以上的纹理参数行差异性检验,绘制受试者工作特征曲线(ROC),计算曲线下面积(AUC)。结果RDA、PCA、LDA、NDA算法的正确率范围分别为:71.67%~91.67%、66.67%~91.67%、68.33%~86.68%、88.33%~93.33%,其中NDA正确率最高。并且MI及FPM提取特征参数的NDA分析法正确率最高,达到93.33%,分类效果最好。腺癌淋巴结转移组的Mean、Perc.01%、Perc.10%、S(0,1)DifEntrp、S(1,-1)DifEntrp大于无淋巴结转移组,仅S(0,4)Correlat值小于无淋巴结转移组,均具有统计学差异。6个纹理参数均具有诊断效能,且S(0,1)DifEntrp诊断效能最好,AUC值达0.840,最佳阈值为1.36,敏感性及特异性分别为92.0%、74.3%。结论基于胸部增强CT纹理分析,MI及FPM特征提取联合NDA分析正确率最高,有助于预测肺腺癌淋巴结转移。
Objective To explore the role of computed tomography(CT)texture analysis in predicting lymph node metastasis of lung adenocarcinoma.Methods The CT enhanced images of 60 cases of lung adenocarcinoma were retrospectively analyzed,including 25 cases with no lymph node metastasis and 35 cases with lymph node metastasis.Region of interest(ROI)was chosen on axial CT images with maximum enhancement of lesion and texture analysis was performed using Mazda software.The feature selection methods included Fishers coefficient,POE+ACC,mutual information(MI),the combination of the above three methods(FPM),and 1-NN,which were used to extract the most significant texture features.Finally,six parameters were extracted that appear three times or more in these five methods.The statistical methods such as RDA,PCA,LDA,and NDA were used and the analysis result was shown in classification Rate.Results Among 304 texture parameters,the classification rate of NDA(88.33%-93.33%)was higher than that of any other kind of method.The NDA of MI and FPM had the highest misclassification rate(93.33%),and the best classification ability.The Mean、Perc.01%,Perc.10%,S(0,1)DifEntrp,S(1,-1)DifEntrp of lymph node metastasis group was significantly higher than those of no lymph node metastasis group,while the S(0,4)Correlat(P<0.05).The AUC of ROC in S(0,1)DifEntrp was 0.84(66.67%sensitivity and 77.78%specificity,P<0.05).However,the AUC of ROC in sigma was 0.621(100.00%sensitivity and 37.04%specificity,P>0.05),which had no statistical significance.Conclusion Based on chest enhancement CT,MI and FPM feature extraction combined with NDA has the highest accuracy,and S(0,1)DifEntrp has the best diagnostic efficiency,which contributes to prediction of lymph node metastasis in lung adenocarcinoma.
作者
徐圆
段钰
曹正业
沈力
王丽娟
叶靖
吴晶涛
XU Yuan;DUAN Yu;CAO Zhengye(Department of Medical Imaging,Northern Jiangsu People’s Hospital Affiliated to Yangzhou University,Yangzhou,Jiangsu Province 225001,P.R.China)
出处
《临床放射学杂志》
CSCD
北大核心
2020年第4期691-695,共5页
Journal of Clinical Radiology
关键词
肺腺癌
纹理分析
体层摄影术
X线计算机
Lung adenocarcinoma
Texture analysis
Tomography,X-ray computed
作者简介
通讯作者:吴晶涛。
