| Objective:To develop a radiomics model of contrast-enhanced CT Imaging to predict postoperative recurrence-free survival(RFS)in patients with CNLC Ⅰ-Ⅱ Stage Hepatocellular Carcinoma(HCC).Materials and methods:A total of 161 eligible patients with CNLC Ⅰ-Ⅱstage HCC were enrolled in this retrospective study.All patients were randomly divided into training(113 patients)and validation(48 patients)sets.Radiomics features were extracted from portal venous phase CT images.Intra-class and inter-class correlation analysis,univariate Cox analysis,least absolute shrinkage and selection operator COX regression model were used for features selection and radiomics score(Rad-score)construction.Univariate and multivariate COX analysis are used to identify clinical risk factors related to RFS.Then,four predictive models are developed using COX regression analysis,including Rad-score,clinical model(containing independent clinical risk factors),CNLC class and combined model(containingindependent clinical risk factors and Rad-score).AUC,C-index,NRI and IDI are used to estimate and compare the discrimination of the four prediction models comprehensively,the calibration curve was used to evaluate the calibration,and the DCA was used to evaluate the clinical utility.Results:11 radiomics features wereselected to establish Rad-score.Sex,tumor pseudocapsule state,and internal neoplastic artery were independent clinical risk factors.The C-index of combined model,Rad-score,clinical model and CNLC class were 0.773(95%CI:0.728-0.818),0.754(95%CI:0.704-0.805),0.743(95%CI:0.676-0.811),0.720(95%CI:0.640-0.801)in the training set,respectively.And the C-index were 0.824(95%CI:0.756-0.893),0.777(95%CI:0.693-0.861),0.802(95%CI:0.685-0.920)and 0.758(95%CI:0.637-0.879)in the validation set,respectively.There were no significant difference in the C-index of the four models(P>0.05)both in training and validation sets.In the training set,the AUC of the combined model in estimating 1,2,3-year RFS were greater than clinical model and CNLC class(P<0.05),and the AUC of Rad-score in predicting 1,2-year RFS were greater than CNLC class(P<0.05),while the AUC of Rad-score in predicting 1,2,3-year RFS were no statistical difference compared with the clinical model and the combined model(P>0.05).In the validation set,the AUC of the combined model in estimating 1,2,3-year RFS were greater than other three models with no statistical significance(P>0.05).The NRI and IDI were greater than 0 with statistically significant both in the training and validation sets when the combined model compared with the clinical model,Rad-score,and CNLC staging(P<0.05).When Rad-score compared with CNLC staging,the NRI and IDI were greater than 0 with statistically significantin the training set(P<0.05),and the NRI and IDI were also greater than 0 in the validation set with only IDI was statistically significant(P<0.05).The NRI and IDI were no statistically significant when Rad-score compared with the clinical model both in the training and validation sets.All four models showed good calibration in the training set,while the calibration of CNLC class was worse in the validation set.The DCA results showed the combined model has the highest clinical net benefit compered with other three models.The net benefit of Rad-score was higher than the clinical model in predicting 1-year RFS,while comparable to the clinical model in predicting 2,3-year RFS.And the net benefit of Rad-score was higher than CNLC class in predicting 1,2,3-year RFS.Conclusion:Rad-score based on portal vein enhanced CT image was a significant predictor of postoperative RFS in patients with CNLC Ⅰ-Ⅱ stage HCC.Combined with clinical risk factors and Rad-score can improve the predictive performance.Radiomics can be used as a supplement to help clinicians screen patients with high risk of recurrence to formulate individualized diagnosis and treatment plans. |
PDF Full Text Request