Predicition Of Early Postoperative Recurrence Of Primary Hepatocellular Carcinoma By Radiomics And Deep Learning Model Based On Gd-EOB-DTPA Enhanced MRI

PART1 Prediction of early postoperative recurrence of prima-ry liver cancer based on clinical risk factorsObjective:To investigate the value of clinical baseline data,MR features,and pathological results in predicting early postoperative recurrence(≤2 years of recurrence)in patients with primary hepatocellular carcinoma(HCC).Materials and Methods:A total of 285 patients with hepatocellular carcinoma confirmed by surgery and pathology and meeting the inclusion criteria of this study from January 2012 to September 2018 in Sun Yat-sen University Cancer Center and Zhujiang Hospital of Southern Medical University were retrospectively collected,and 70%of the cases(195 cases)were used to train the model and 30%of the cases(90 cases)were used to validate the model.They were divided into early recurrence(≤2 years recurrence,ER)and non-early recurrence groups according to the follow-up results.Mann-Whitney test,chi-square test and Fisher’s exact test were used to compare age,gender,neutrophils count(NE),hepatitis B DNA virus quantification,alanine aminotransferase(ALT),aspartate aminotransferase(AST),y-glutamyl transpeptidase(GGT),alpha-fetoprotein(AFP),tumor size,arterial phase enhancement pattern,internal arteries,capsule appearance,hypodense halo,intratumoral necrosis,satellite nodules,peritumoral hypointense,hepatobiliary specific phase intensity(HBP intensity),microvascular invasion(MVI),tumor number,and BCLC stage between the two groups.Variables with statistical significance after univariate analysis were included in multivariate logistic regression analysis to construct a model which was visualized as a nomogram based on imaging features and clinical factors,the receiver operating characteristic(ROC)curve and area under the curve(AUC)was used to assess the diagnostic efficacy of the model.Results:Univariate analysis in the training set showed that NE,AST,GGT,AFP,internal arteries,HBP intensity,MVI,tumor number,and BCLC stage(P<0.1)were risk factors for early postoperative recurrence in patients with HCC,and multivariate logistic regression analysis showed that NE,AST,MVI,and HBP intensity(P<0.05)were independent risk factors for predicting early postoperative recurrence in patients with HCC,and we build a predictive nomogram,the AUC of the ROC curve analysis model was 0.761(95%CI:0.684-0.837)(training set),0.720(95%CI:0.594-0.846)(validation set),respectively.Conclusion:The imaging features based on Gd-EOB-DTPA enhanced MR,clinical and pathological risk factors can predict the ER of patients with postoperative HCC.PART2 Preliminary study of radiomics model based on Gd-EOB-DTPA enhanced MR to predict early postoperative re-currence of primary hepatocellular carcinomaObjective:To investigate the feasibility of radiomics signature based on Gd-EOB-DTPA enhanced MRI in predicting ER of postoperative HCC,and to compare the predictive efficacy of the above model with the combined clinical-radiomics model,and select the best model for predicting early recurrence of postoperative HCC.Materials and Methods:The clinical baseline data,MR imaging features,pathological results,follow-up protocol,MR scan method of the cases included in this chapter are the same as those in chapter 1.Based on the whole tumor region of interest(ROI)segmentation and feature extraction of HCC in arterial phase(AP),portal phase(PVP)and hepatobiliary specific phase(HBP),after dimensionality reduction and feature selection in AP,PVP,HBP,the radiomics signature of three phases individually and the 3-phases combined radiomics signature were built respectively.ROC curve was used to evaluate the diagnostic efficacy of each model,select the optimal model,combined with the clinical factors related to ER summarized in chapter 1,and then established a clinic-radiomics prediction model,evaluate the diagnostic efficacy of the model in the training set and validation set,respectively.Results:The results showed that the radiomics signature composed of 8 features in AP,17 in PVP and 13 in HBP was associated with ER of postoperative HCC(P<0.000),and the 3-phases combined radiomics signature was also associated with ER of HCC(P<0.000).The diagnostic efficacy of the three-phase separate model.The AUC of AP radiomics signature was 0.759(95%CI:0.690-0.828)(training set)and 0.727(95%CI:0.641-0.813)(validation set).The AUC of PVP radiomics signature was 0.678(95%CI:0.594-0.762)(training set)and 0.689(95%CI:0.599-0.779)(validation set),respectively.The AUC of HBP was 0.676(95%CI:0.586-0.754)(training set)and 0.702(95%CI:0.614-0.790)(validation set),respectively.The diagnostic efficacy of the 3-phases combined radiomics signature obtained good predictive performance:0.759(95%CI:0.686-0.833)(training set),0.729(95%CI:0.608-0.851)(validation set),while the model predictive power of the clinical-radiomics prediction model was 0.830(95%CI:)(training set),0.748(95%CI:)(validation set),respectively.The results are higher than the four imaging models established in this chapter and also higher than the traditional clinical prediction models established in chapter 1.Conclusion:Prediction model based on Gd-EOB-DTPA enhanced MRI could be helpful for predicting ER of postoperative HCC.PART3 Preliminary study of deep learning model based on Gd-EOB-DTPA enhanced MR to predict early postoperative recurrence of primary hepatocellular carcinomaObjective:The purpose of this part is to investigate the feasibility of deep learning signature based on Gd-EOB-DTPA enhanced MR to predict ER of postoperative HCC,and to compare the efficacy of deep learning signature,clinical nomogram and the clinical-radiomics model.Thus,we aimed to create a optimal model for predicting ER of postoperative HCC among all the prediction models in this study.Materials and Methods:The clinical baseline data,MR imaging features,pathological results,follow-up protocol,MR scan method and the whole tumor ROI delineation of the cases included in this chapter are the same as those in chapter 1 or 2.Respectively establish the deep learning model of AP,PVP,HBP based on VGGNet19 network and the deep learning model of combined 3-phases,use the ROC curve to evaluate the diagnostic efficacy of each model and select the optimal model,combine the clinical factors related to ER of postoperative HCC summarized in chapter 1,establish the clinical-deep learning prediction model,the performance of different models(involving the clinical model and clinical-radiomics model)were analyzed by AUC,calibration curve and decision curve.Results:The deep learning model was established to predict ER of HCC based on AP,PVP and HBP,respectively.The results showed that the model achieved good predictive power in each phase in the training set and validation set.The AUC of arterial phase model was 0.882(95%CI:0.823-0.941)(training set)and 0.826(95%CI:0.755-0.897)(validation set),the AUC of portal phase model was 0.822(95%CI:0.753-0.891)(training set)and 0.854(95%CI:0.780-0.928)(validation set),and the AUC of hepatobiliary specific phase model was 0.909(95%CI:0.860-0.958)(training set)and 0.888(95%CI:0.833-0.943)(validation set),respectively.However,the predictive power of the established clinical-deep learning model was 0.949(95%CI:0.919-0.980)(training set)and 0.908(95%CI:0.841-0.975)(validation set),respectively,and the results were higher than those of the clinical-imaging model established in the previous chapter and also higher than those of the traditional clinical prediction model established in chapter 1.The results of calibration curve analysis also showed that the nomogram of the clinical-deep learning model had better predictive accuracy,and decision curve analysis showed that this model could achieve higher net benefits for patients.Conclusion:In this study,we successfully established a deep learning model based on Gd-EOB-DTPA enhanced MR to predict early postoperative recurrence of HCC,and further established a clinical-deep learning prediction model,which can more accurately predict early postoperative recurrence in HCC patients,and improve the net benefit of patients and help guide clinical decisions.