Research On CT-based Radiomcis For Predicting TACE Refractoriness In Hepatocellular Carcinoma

[Objective]To identify CT-based radiomics model for individually prediction of the risk of TACE refractoriness in HCC before treatment.[Materials and methods]We retrospectively collected 160 patients with HCC(97 cases of TACE refractoriness and 63 cases without TACE refractoriness)of HCC in our hospital from October 2013 to October 2020,and analyzed their preoperative CT image characteristics and clinical data.It is used to manually sketch lesions on CT in non-contrast phase(NP),arterial phase(AP),vein phase(VP),and delayed phase(DP)CT layer by layer,forming a 3D volume of interest(VOI).Based on the PyRadiomics package,1218 radiomics features were extracted from each phase.Part Ⅰ:Based on four single-phase features and one full-phase feature,two independent samples t-test、mRMRandLassoCV were used to select features,then use Logistic regression to build prediction models separately;Logistic,SVM,RFC model to be developed.The model evaluation used ROC,AUC,sensitivity,specificity,accuracy,and DeLong test to compare the difference between AUC.Part Ⅱ:Based on the clinical data and conventional radiological characteristics of HCC patients,the independent risk factors associated with TACE refractoriness are analyzed through univariate and multivariate Logistic regression,and then the clinical radiological model is builded using Logistic regression.The model predicting the best performance in the first part as the radiomics model,and a combined model is estabilised by combining the Radscore clinical radiological independent factors.At the same time,the combined model is visualized in nomogram.In addition to the above evaluation indicators,the consistency of the predictive models was evaluated by the Hoster-Lemeshow goodness-of-fit test and calibration curves,decision curves are used to assess that the clinical benefit of combined model for HCC patients.[Results]Part Ⅰ:The prediction performance of single-phase and full-phase radiomic models and full-phase different machine learning models are compared.1.The full-stage phase model has the best performance,and the AUC in the training set and the test set are 0.850 and 0.800 respectively,but the AUC are only statistically different from the DP model in the test set(P<0.039);and there is no statistical significance between the other models in training set(P>0.05).2.SVM has the best prediction performance,and the AUC value in the test set is 0.826(95%CI:0.691-0.953),which is higher than that of the RFC and LR models(0.822,0.800),but the difference is statistically significant(P<0.05).Part Ⅱ:In clinical radiological features,AFP levels and the number of lesions are independent risk factors in predicting TACE refractoriness.In the training set,the predictive performance of the combined model for TACE refractoriness is significantly better than that of the clinical model(AUC:0.919 vs 0.632)and slightly higher than that of the radiomics model(AUC:0.900);In the test set,the combined model was significantly higher than the clinical model(AUC:0.851 vs 0.639),and the difference is statistically significant(P=0.043),the performance of combined model is better than that of the radiomics model(AUC:0.826).Based on Youden index,the optimal cutoff of the combined model is 0.467,HCC patients are divided into a high-risk group(risk value>0.467)and a low-risk group(risk value≤0.467).[Conclusion]The combined model using clinical variables、conventional radiological characteristics and radiomic features can predict TACE refractoriness in HCC.The nomogram can assess the risk of TACE refractorinss as a visual tool,provide a reference for clinical decision.