Construction And Application Of Prognostic Models For Perihilar Cholangiocarcinoma After Radical Resection Based On Magnetic Resonance Imaging Radiomics And Clinical Factors

Objective:Perihilar cholangiocarcinoma（pCCA）is the most common biliary malignant tumor.The key point of individualized treatment of pCCA is to accurately predict survival probability before radical resection and early recurrence（ER）after resection,and to stratify patients before planning treatment and follow-up.In this study,magnetic resonance imaging based radiomics combined with clinical factors was used to construct models for predicting overall survival and early recurrence of pCCA after radical resection.Materials and Methods:1.A total of 184 patients with pCCA underwent radical resection and had complete clinicopathological information in West China Hospital of Sichuan University were retrospectively collected.All of the patients were confirmed by pathology.The patients were randomly divided into training group（n=110）and testing group（n=74）.Univariate and multivariate Cox regression analysis were used to screen the potential clinical risk factors related to overall survival（OS）.The independent risk factors related to OS were used for the construction of clinical model.A comprehensive prediction model was constructed by combining clinical risk factors and the arterial-and portal-phase radiomic features extracted from contrast-enhanced magnetic resonance images.The constructed models were visualized as Nomogram.The prediction performance of the models was evaluated by the Harrell’s Concordance Index（C-index）,calibration curve and decision curve analysis.The Kaplan-Meier curve was used to describe survival rates in patients with high and low risk stratification,and log-rank test was used for their comparison.2.Patients were randomly divided into training group（n=128）and testing group（n=56）.The potential clinical risk factors related to postoperative early recurrence were screened by univariate and multivariate logistic regression analysis.The risk factors related to ER were obtained and used to construct the clinical model.A comprehensive prediction model was constructed by combination of radiomic features and clinical risk factors.The constructed models were visualized as Nomogram.The prediction performance of the model was evaluated by the receiver operating characteristic curve（ROC）,calibration curve and decision curve analysis.Kaplan-Meier curve was applied to illustrate the survival rate of patients with high and low risk of ER stratified by model,and log-rank test was used for their comparison.Results:1.Multivariate Cox regression analysis showed that preoperative CEA（HR=3.16,95%CI:1.44-6.94,P=0.004）,_CN stage（HR=1.76,95%CI:1.08-2.87,P=0.023）and hepatic artery invasion in images（HR=4.11,95%CI:1.96-8.64,P<0.001）were independent risk factors for poor prognosis.In the training group,the model combining radiomics signature(Signature^AP and Signature^PVP)and clinically relevant risk factors had the best performance in predicting the overall survival,with a C-index of 0.962（95%CI:0.905-1）,which was better than that of the sole clinical model（0.846,95%CI:0.735-0.957）,Signature^AP（0.871,95%CI:0.771-0.971）,Signature^PVP（0.709,95%CI:0.504-0.914）and the 8^th edition of AJCC staging system（0.616,95%CI:0.522-0.711）.In the testing group,the comprehensive model also had the best performance in predicting survival,with a C-index of 0.814（95%CI:0.569-1）,which was better than that of the clinical model（0.755,95%CI:0.540-0.969）,Signature^AP（0.774,95%CI:0.637-0.910）,Signature^PVP（0.678,95%CI:0.485-0.871）,and the 8^thedition of AJCC staging system（0.599,95%CI:0.491～0.708）.Calibration curve and clinical decision curve analysis also confirmed that the comprehensive model had the highest prediction accuracy and the maximum clinical net benefit.Kaplan-Meier curve showed that there was a statistically significant difference in OS between the high and low risk group stratified by the comprehensive model in training group（P<0.001）,and there was also a statistically significant difference in OS in the testing group（P=0.025）.2.Multivariate Logistic regression analysis showed that preoperative blood glucose level（OR=0.265,95%CI:0.071-0.868,P=0.036）,white blood cell count（OR=0.265,95%CI:0.071-0.868,P=0.001）,CEA（OR=2.962,95%CI:1.166-8.014,P=0.026）and hepatic artery invasion in imaging（OR=6.134,95%CI:2.113～20.833,P=0.002）were independent related factors for ER.In the training group,the comprehensive model had the best performance in predicting ER,with an AUC of 0.868（95%CI:0.807-0.928）,which was better than the clinical model（0.841,95%CI:0.770-0.912）,Signature^AP（0.664,95%CI:0.570-0.759）and the 8^thedition of AJCC staging system（0.619,95%CI:0.619）.The specificity,sensitivity,accuracy,negative predictive value and positive predictive value of the comprehensive model were 0.736,0.857,0.789,0.869 and 0.716,respectively.In the testing group,the comprehensive model also had the best performance in predicting ER,with an AUC of 0.835（95%CI:0.727-0.943）,which was better than that of the clinical model（0.821,95%CI:0.712-0.931）,Signature^AP（0.663,95%CI:0.519-0.807）,and the 8^th edition of AJCC staging system（0.574,95%CI:0.427-0.722）.The specificity,sensitivity,accuracy,negative predictive value and positive predictive value of the comprehensive model were 0.710,0.760,0.732,0.786 and 0.679,respectively.The calibration curves showed a good agreement between the actual early recurrence probability and the early recurrence probability predicted by the clinical model and the comprehensive model.The decision curve analysis showed that the clinical model and the comprehensive model were superior to the Signature^AP and the 8^th edition of AJCC staging system in the training and testing groups.Kaplan-Meier curves also showed a significant difference in OS between high and low risk group of ER stratified by the comprehensive model in the training and testing groups（both,P<0.001）.Conclusion:The model combining radiomics and clinical parameters can be used to evaluate the prognosis of pCCA patients.Radiomics Signature is a potential quantitative tumor marker for prognosis.It is a non-invasive,low-cost tool for preoperative predicting survival and ER in patients with pCCA,which could be used to help develop follow-up plan and make clinical treatment strategies.As an important supplement to precision medicine,it has positive significance for the realization of prolonging the postoperative survival of pCCA patients and the implementation of individualized treatment.