CT图像纹理分析鉴别不典型胰腺实性假乳头状肿瘤与胰腺导管腺癌的初步研究

Texture analysis based on CT images to differentiate atypical pancreatic solid pseudopapillary tumor and pancreatic adenocarcinoma:a preliminary study

目的探讨采用CT图像纹理分析鉴别不典型的胰腺实性假乳头状肿瘤(pancreatic solid pseudopapillary tumor,SPT)和胰腺导管腺癌(pancreatic ductal adenocarcinoma,PDAC)的可行性。方法回顾性分析四川大学华西医院经病理学检查证实的不典型SPT(共计26个病灶)和PDAC(共52个病灶)患者的CT资料。利用ITK-Snap软件于动脉期(arterial phase,AP)及门静脉期(portal venous phase,PVP)CT图像上勾画三维(three-dimensional,3D)感兴趣区(region of interest,ROI),利用A.K.软件(GE公司,美国)自动提取ROI处的图像纹理特征。应用R软件行参数间的相关性分析以去除冗余的纹理特征后,剩余的纹理特征应用单因素及多因素二分类logistic回归筛选纹理特征,并建立回归模型。应用受试者工作特征(receiver operating characteristic,ROC)曲线分析比较纹理特征与模型鉴别不典型SPT及PDAC的诊断效能。结果共提取了792个纹理特征(AP396个,PVP 396个),去冗余后剩余61个特征(AP 35个,PVP 26个)。二分类logistic回归分析选择出2个纹理特征为独立危险因素(AP下为MinIntensity,PVP下为Correlation_AllDirection_offset1_SD),其鉴别不典型的SPT和PDAC的灵敏度、特异度分别为71.15%、76.92%和63.46%、76.92%,曲线下面积(AUC)分别为0.740和0.754。应用上述2个纹理特征建立二分类logistic模型后,其模型灵敏度和特异度分别为73.08%及80.77%,AUC值为0.796。2个纹理特征和logistic模型的诊断效能比较差异无统计学意义(P>0.05)。结论 CT图像纹理分析鉴别不典型SPT与PDAC是可行的,具有中等的诊断效能。

Objective To access the diagnostic performance of CT texture analysis to differentiate atypical pancreatic solid pseudopapillary tumor（SPT） from pancreatic ductal adenocarcinoma（PDAC）. Methods CT images of26 patients with pathologically proved atypical SPT and 52 patients with PDAC were analyzed. 3D regions of interest（ROIs） on arterial phase（AP） and portal venous phase（PVP） images were drawn by ITK-Snap software. A.K. software（GE company, USA） was used to extract texture features for the discrimination of atypical SPT and PDAC. After removing redundancy（by a correlation analysis through R software）, texture features were selected by single-factor and multi-factor logistic regression, and logistic regression model was finally established. Receiver operating characteristic（ROC） analysis was performed to assess the diagnostic performance of single texture feature and logistic model. Results A total of 792 texture features [396 of AP, 396 of PVP] from AP and PVP CT images were obtained by A.K. software. Of these, 61 texture features（35 of AP, 26 of PVP） were selected by R software（result of correlation analysis showed that correlation coefficient 〉0.7）. Two texture features, including MinIntensity and Correlation_AllDirection_offset1_SD, were selected to establish logistic model. The sensitivity and specificity of these 2 texture features were 71.15% and 76.92%, 63.46% and76.92% respectively, the area under curve（AUC） were 0.740 and 0.754 respectively. The model＇s sensitivity and specificity were 73.08% and 80.77% respectively, the AUC value was 0.796. There was no significance among the model,MinIntensity, and Correlation_AllDirection_offset1_SD（P〉0.05）. Conclusions CT texture analysis of 3D ROI is a quantitative method for differential diagnosis of atypical SPT from PDAC.