| Purpose:To construct a comprehensive prediction model of progression free survival time of pancreatic cancer after curative operation based on CT radiomics analysisMaterials and Methods:A retrospective analysis of 162 patients with pancreatic cancer who had undergone curative surgery and enhanced CT examination before operation and were followed up to disease progression or for at least 2 years.According to the follow-up results,the patients were divided into disease recurrence group and non-disease recurrence group.Kaplan Meier curve was used for univariate analysis to select clinical-radiological parameters.Variables with P value less than 0.1 in univariate analysis were included in Cox proportional risk model.Forward method was used for multivariate analysis to select independent prognostic factors and establish the clinical-radiological prediction model.The tumor was segmented on three phase images(arterial phase,portal phase and delayed phase)of preoperative enhanced CT to extract the radiomics features.Feature correlation analysis and Bayesian model averaging(BM A)were used to screen key radiomics features.Random survival forest(RSF)was used to establish radiomics models and the combination model.Five-fold cross validation was used for internal validation.Based on the prediction risk of each RSF model,the integrated area under the time-dependent ROC curve(iAUC)was used to calculate and compare the prediction performance of the models.Results:The clinical parameters with independent prognostic value were CA199 level(cutoff value is 37U/ml)and tumor differentiation(poor differentiation vs medium-well differentiation).The prediction ability(iAUC)of clinical-radiological model,arterial phase,portal phase,delayed phase and three-phase combined radiomics features were 0.696(95%CI:0.632,0.735),0.734(95%CI:0.686,0.778),0.779(95%CI:0.736,0.822),0.733(95%CI:0.736,0.822),0.741(95%CI:0.694,0.790),respectively.In the final combined model,the independent risk factors for predicting the progress of postoperative diseases were portal phase radiomics features(HR=2.945,95%CI:1.980,4.380,P<0.001),postoperative CA199 higher than 37U/ml(HR=1.596,95%CI:1.087,2.344,P=0.017),and poor differentiation(HR=1.525,95%CI:1.021,2.276,P=0.039)and the iAUC was 0.788(95%CI:0.745,0.832),which was statistically higher than that with the clinical model(P<0.001).Conclusion:The combination of the three-phase radiomics features,CA199 higher than 37U/ml and poor tumor differentiation has a high clinical value in predicting the recurrence of pancreatic cancer after curative resection,but its clinical value still needs further external validation.Purpose:To construct a comprehensive prediction model for early recurrence(12 months)of pancreatic cancer after curative surgery based on CT radiomics.Materials and methods:A total of 164 patients with pancreatic cancer who had undergone resection and enhanced CT examination before operation and were followed up for at least 1 years.According to the follow-up results,the patients were divided into early recurrence group and non-early recurrence group.Univariate and multivariate logistic regression analysis was applied to screen valuable clinical-radiological parameters and build clinical-radiological prediction model.Dedicated radiomics prototype software(radiomics)was used to segment tumor on preoperative enhanced CT including arterial phase,portal phase and delayed phase.The radiomics features were extracted.The correlation analysis and Gini index were used to screen the useful radiomics features.The four radiomics models(based on arterial phase,portal phase,delay phase and three phase combination)was built using logistics regression analysis.The best radiomics model was selected according to the area under the curve(AUC)of receiver operator characteristic curve.Then radiomics signature and clinical-radiological features were combined to build the combined model.The operating characteristic curves(ROCs)were used to evaluate the differential diagnostic performance of the three models.Nomogram was used to present the final model with best performance.C-index was used to evaluate the identification ability of nomograph,and calibration curve was used to evaluate the accuracy of nomograph.The clinical value of nomogram was evaluated by decision curve analysis(DCA).Results:In multivariate analysis,only post-operative CA199 value(cutoff value 185U/ml)was the independent predictor for early recurrence and the AUC value was 0.701(95%CI:0.617,0.785).The area under curve(AUC)for predicting early recurrence of pancreatic ductal adenocarcinoma after radical resection was 0.644(95%CI:0.572,0.731)with arterial phase radiomics signature,0.668(95%CI:0.601,0.734)with portal phase radiomics signature,0.604(95%CI:0.534,0.689)with delayed phase radiomics signature,0.722(95%CI:0.664,0.818)with three-phase combined radiomics signature,respectively.The combination of three-phase combined radiomics signature and post-operative CA199 value(cutoff value 185U/ml)showed the best prediction performance and the AUC value was 0.812(95%CI:0.746,0.877.The nomogram based on the combined model achieve good calibration ability(Brier score was 0.179).Conclusion:The nomogram combined of three-phase radiomics score and postoperative CA199 higher than 185U/ml show good performance in predicting early recurrence of pancreatic ductal adenocarcinoma after curative resection,but its clinical value still needs further external validation.Purpose:to investigate the performance of CT radiomics for distinguishing post-operative changes from local recurrence in patients with pancreatic cancer after surgical resectionMaterials and Methods:A retrospective review of 153 patients who underwent surgical resection of pancreatic cancer with regular post-operative follow-up were consecutively enrolled and divided into local-recurrence(n=83)and post-operative change(n=70)group based on biopsy or clinical follow up results.Univariate and multivariate logistic regression analysis was used to select valuable clinical and traditional imaging features.Dedicated radiomics prototype software was used to segment postoperative local-regional soft tissue on portal vein phase CT images and extract radiomics features.The Gini index of random forest was used to select useful radiomics features and the binary logistic regression was used to establish the radiomics model.Then radiomics signature and clinical-radiological features were combined to build the combined model.The operating characteristic curves(ROCs)were used to evaluate the differential diagnostic performance of the three models.Nomogram was used to present the final model with best performance.C-index was used to evaluate the identification ability of nomograph,and calibration curve(Brier score)was used to evaluate the accuracy of nomograph.The clinical value of nomogram was evaluated by decision curve analysis(DCA).Results:For clinical-radiological model,the useful parameters for differential diagnosis were portal vein stricture and the CT value of the soft tissue on delayed phase and the diagnostic performance of the clinic-radiological model was 0.640(95%CI:0.553,0.730)The radiomics model achieved favorable discriminatory ability with an area under the curve(AUC)of 0.750(95%CI,0.675 to 0.830).The AUC of the combined model incorporating the above parameters was 0.775(95%CI:0.699,0.850).The nomogram based on the combined model achieve good calibration ability(Brier score was 0.191).Conclusion:The combination of the radiomics score,the delayed enhancement CT value of the soft tissue in the operation area and the local vascular stenosis reached a high clinical value in the diagnosis of local recurrence of the soft tissue in the operation area,but its clinical value still needs further external validation. |
PDF Full Text Request