Differentiating Hypovascular Pancreatic Neuroendocrine Tumor From Pancreatic Ductal Adenocarcinoma Based On CT Features And Texture Analysis

Purpose Hypovascular pancreatic neuroendocrine tumors(hypo-PNETs)have similar imaging findings with pancreatic ductal adenocarcinomas(PDACs),that were easily misdiagnosed as PDACs.The aim of the current study is to investigate the values of computed tomography(CT)features and CT texture analysis in differentiating hypovascular PNET from PDAC and to improve the ability of diagnosis and differential diagnosis.Materials and Methods 1.Subjects 155 patients with PDACs between September 2012 and August 2017 and 62 patients with PNETs between January 2010 and June 2017 who confirmed by pathology were collected.According to the exclusion criterias,49 patients with PNETs and 79 patients with PDACs were excluded,we randomly selected 39 patients with PDACs at the ratio of 1: 3 from the remaining 76 PDACs.Finally 13 patients with hypo-PNETs and 39 patients with PDACs were included in this retrospective study.All these patients underwent preoperative unenhanced and dynamic contrast-enhanced CT examinations.2.CT image analysis The author and another abdominal radiologist independently reviewed the CT images.The CT imaging findings include the following parameters: the tumor’s size(the maximum diameter),tumor location(head or neck,and body or tail),margin(well-defined,ill-defined),mean CT attenuation value(unenhanced,pancreatic parenchymal phase,portal venous phase),calcification,distal pancreatic parenchymal artrophy,pancreatic duct dilatation,intra-/extrahepatic bile duct dilatation,lymphnodes invasion,local invasion or distant metastases.3.CT texture analysis All CT images were exported in DICOM format from the picture archiving and communication system(PACS)for further texture analysis.The tumors showed relatively clear on pancreatic parenchymal phase,and were easily to draw,so we firstly imported the pancreatic parenchymal phase images into GE Omni-Kinetics software.For each lesion,we selected three axial images as region of interests(ROIs).The first axial image was obtained at the largest cross-section image,the second and the third images were obtained on the previous and the next image based on the largest axial image,respectively.The two radiologists draw the ROIs along the margin of the tumor,independently.After the tumors were segmented,the texture features were automatically extracted by the software.The above procedures were repeated on unenhanced and portal venous phase images.A total of 68 texture features were extracted on per phase,and they were separated into four categories:(1)First Order Histogram statistics;(3)Grey level cooccurrence matrix(GLCM);(4)Haralick;(5)run length matrix(RLM).4.Statistic analysis All statistic analyses were performed using SPSS Version 20.0.The CT imaging findings between hypo-PNETs and PDACs were compared using χ2 test or Fisher exact test for categorical variables and the dependent sample t-test for continuous variables,then for the CT imaging findings with significant statistic differences,ROC and multivariate logistic regression analysis was performed to select the important different-iators.For the texture features with normally distributed variables,the dependent sample t-test was used to identify the differences between hypo-PNETs and PDACs.The MannWhitney U test was used for these un-normal distributed parameters.For the texture features with statistic difference,univariate regression analysis was performed,for the texture features with statistic differences at univariate regression analysis,receiver operating characteristics(ROC)curve analysis was performed and to calculate the AUC,sensitivity and specificity.Logistic regression analysis was performed to select the significant differenttiators of hypo-PNETs from PDACs.P<0.05 was considered to indicate a significant difference.Results 1.For CT image analysis,there were significant statistic differences in margin,pancreatic duct dilatation and local invasion/ metastasis between hypo-PNETs and PDACs(P=0.01,0.016,0.006,respectively).The AUCs were 0.718,0.718,0.705,respectively.The tumor’s size,location,CT attenuation value per phase,calcification,distal pancreatic parenchymal artrophy and intra-/extrahepatic bile duct dilatation have no statistic differentces between the two entities.After logistic regression analysis,pancreatic duct dilatation was independent differentiators of hypo-PNETs from PDACs,the AUC was 0.718,sensitivity of 0.590 and specificity of 0.846.2.On unenhanced texture analysis,Median Intensity,std Deviation,Range,Mean Deviation,Entropy,uniformity,MPP,Quantile5/10/25/50,skewness and Grey Level Nonuniformity,Haralick correlation and variance have significant differences between hypo-PNETs and PDACs.After logistic regression analysis,entropy(OR=0.024,95% confidence interval(95%CI)=0.001-1.81),Quantile5(OR=0.734,95%CI=0.58-0.93)were independent differentiators between the two kinds of tumors,the AUC for the diagnostic model was 0.846,sensitivity of 0.795 and specificity of 0.846.3.On pancreatic parenchymal phase texture analysis,Median Intensity,RMS,MPP,Mean Deviation,uniformity,Quantile5/10/25/50/75/90/95,Voxel Value Sum and GLCM total frequency have significant differences between the two tumors.After logistic regression analysis,MPP(OR=0.082,95%CI=0.01-0.75),Quantile25(OR=5.232,95%CI=1.11-24.69)and Quantile95(OR=2.102,95%CI=1.09-4.05)were valuable for differentiating hypo-PNETs from PDACs,the AUC for the diagnostic model was 0.870,with 0.846 sensitivity and 0.923 specificity.4.On portal venous phase texture analysis,there were significant differences in contrast,sum Average,Difference Variance and Grey Level Nonuniformity between hypoPNETs and PDACs.After logistic regression analysis,contrast(OR=6.08,95%CI=1.17-31.66)was independent differentiator hypo-PNETs from PDACs,the AUC for the diagnostic model was 0.716,sensitivity of 0.436 and specificity of 1.00.5.The significant texture features that were extracted on multivariate regression analysis based on unenhanced,pancreatic parenchymal phase and portal phase images were selected as input variables for multivariate regression analysis to establish the model for texture analysis,the AUC was 0.929,with 0.846 sensitivity and 0.923 specificity,significantly higher than CT imaging diagnostic model(AUC=0.718).Conclusion 1.The significant CT features for differentiating the two entities include margin,pancreatic duct dilatation and local invasion/metastasis.Especially pancreatic duct dilatation can be an independent CT predictor for hypo-PNETs.There were no statistic differences between the two entities in tumor’s size,location,CT attenuation value per phase,calcification,distal pancreatic parenchymal artrophy and intra-/extrahepatic bile duct dilatation.2.The significant CT texture feature parameters for differentiating the two entities includes Entropy and Quantile5 on unenhanced TA,MPP、Quantile25 and Quantile95 on pancreatic parenchymal phase TA,contrast on portal phase TA.When we combined the six significant texture features,the diagnostic performance,sensitivity,specificity were very high.3.The diagnostic performance of texture analysis is superior to that of CT image analysis,and texture analysis can help radiologists to improve the differentiation of hypo-PNETs from PDACs.