A Nomogram Based On CT Radiomics For Preoperative Prediction Of Microvascular Invasion In Hepatocellular Carcinoma

Microvascular invasion(MVI),as a histopathological feature,is one of the important factors affecting the prognosis of hepatocellular carcinoma(HCC)patients and the selection of personalized treatment strategies.However,it is still a challenge to accurately assess the presence or absence of MVI before surgery.In recent years,with the development of artificial intelligence and big data processing technology,radiomics,as an emerging image analysis technology,has attracted more and more attention in the field of tumor research.Compared with traditional radiological image features,radiomics can quantitatively analyze the changes in the subtle structure of the tumor on the image,objectively reflect the tumor heterogeneity,and also provide the possibility for the assessment of tumor biological characteristics.Based on this,this study aims to construct a non-invasive,highly accurate and repeatable model for preoperative individualized prediction of MVI status by comprehensively analyzing the patient’s preoperative CT radiomics,radiological features and clinical data.Part Ⅰ Prediction of microvascular invasion in hepatocellular carcinoma based on clinical or radiological featuresAims:To evaluate the potential value for prediction of MVI status in HCC patients by comprehensively analyzing the preoperative clinical data or radiological features.Methods:We retrospectively collected 381 patients who had undergone preoperative Dynamic Contrast-Enhanced CT(CECT)at our institution from January 2013 to December 2018 and were pathologically confirmed HCC after surgery.The cases were divided into training set(300 cases)and validation set(81 cases)according to the operation date at a ratio of 4:1.Meanwhile,two independent follow-up cohorts(144 cases in our center and 293 cases of HCC patients in the TCGA database)were enrolled to evaluate the impact of MVI characteristic on the prognosis and survival of patients.The patient’s clinical data were the results of laboratory examinations within one week before the operation,and the CT radiological features were independently evaluated by two professional radiologists.Through univariate and multivariate logistic regression analysis of clinical features(CF)or radiological features(RF)in the training dataset,a CF model and RF model for MVI classification were constructed respectively.Sensitivity,specificity,accuracy,receiver operating characteristics curve(ROC)and area under curve(AUC)were used to evaluate the classification performance of models for MVI in the training and validation datasets,respectively,and the Delong test was used to evaluate the difference.The decision curve analysis(DCA)was used to evaluate and compare the clinical applicability of predictive models.Results:1.Survival analysis based on two independent follow-up cohorts showed that MVI positive was an important risk factor for tumor recurrence-free survival in HCC patients,but the impact on patients’ overall survival was different.2.The analysis based on the baseline data of the cases showed that there was no statistical difference in the distribution of the clinical and radiological features between the two cohorts(p>0.05).3.Preoperative clinical features:HBV-DNA quantification,albumin to globulin ratio,alkaline phosphatase,total bilirubin,alpha-fetoprotein and tumor diameter could be served as independent predictors of MVI.4.For MVI classification,the sensitivity,specificity,accuracy,and AUC values of the CF model in the training and validation datasets were 75.2%and 63.3%,72.3%and 68.6%,73.3%and 66.7%,80.6%(95%CI,75.6-85.6%)and 75.9%(95%CI,65.6-86.3%),respectively.5.The radiological features of liver cirrhosis,tumor growth pattern,pseudo-capsule,satellite nodule,arterial rim enhancement and radiogenomics venous invasion could be served as independent risk factors for MVI prediction.6.For MVI classification,the sensitivity,specificity,accuracy,and AUC values of the RF model in the training and validation datasets were 74.3%and 73.3%,85.9%and 86.3%,81.7%and 81.5%,87.2%(95%CI,83.1-91.4%)and 82.0%(95%CI,71.8-92.2%),respectively.7.The RF model was better than the CF model in predicting MVI in the two cohorts,and it had higher clinical applicability.Conclusion:MVI is an important risk factor for the prognosis of HCC patients.Models constructed based on clinical or radiological features provide the possibility to assess the MVI status of HCC patients before surgery,and the predictive value of radiological features is higher.Part Ⅱ Prediction of microvascular invasion in hepatocellular carcinoma based on multi-sequence CT radiomics analysisAims:Radiomics can quantitatively characterize the changes in the subtle structure of tumors in medical images,and objectively reflect tumor heterogeneity.This study continues to explore the significance of MVI prediction based on multi-sequence dynamics-enhanced CT radiomics features of tumors.And by constructing and verifying the effectiveness and clinical value of a nomogram based on radiomics fusion clinical-radiological features for predicting the MVI status of HCC patients before surgery.Methods:The enrolled cohort and dataset distribution were consistent with the first part of the study.The ITK-SNAP software was used to delineate the original boundary of tumors in the arterial and portal-venous phase CT images slice by slice to obtain the 3D volume of interest(VOI).The image expansion and contraction algorithms were used to process the original VOI of the tumor to obtain the corresponding tumor dilated VOI and narrow VOI,respectively.The Pyradiomics software package based on Python was used to extract 1,218 radiomics features for each tumor VOI.The consistency assessment and dimensionality reduction of radiomics features were performed by the intra-class correlation coefficient and least absolute shrinkage and selection operator(LASSO),respectively.And the corresponding radscores were obtained.A nomogram for MVI prediction was constructed by regression analysis of Radscore,clinical and radiological features in the training dataset.Sensitivity,specificity,accuracy,ROC and AUC values were used to evaluate the classification performance of models for MVI in the training and validation datasets,respectively,and the Delong test was used to evaluate the difference.Calibration curve and DCA were used to evaluate the goodness of fit and clinical application of the predictive models,respectively.Results:1.In the radiomics analysis,six radscores constructed by different combinations were of certain value for MVI classification in the two cohorts(AUC=77.4-90.4%,p<0.05),indicating the feasibility of tumor radiomics features for predicting MVI status.2.The PP_Radscoreornginal had the best classification performance for MVI,with the AUC values of 90.4%and 86.4%in the training and validation datasets,respectively.3.In the nomogram,albumin to globulin ratio,aspartate aminotransferase,total bilirubin,neutrophil-lymphocyte ratio,liver cirrhosis,tumor growth pattern,pseudo-capsule,arterial rim enhancement and PP_Radscoreornginal were independent predictors of MVI.4.For MVI classification,the sensitivity,specificity,accuracy,and AUC values of the nomogram in the training and validation datasets were 96.3%and 80.0%,82.2%and 84.3%,87.3%and 82.7%,95.8%(95%CI,93.8-97.8%)and 91.2%(95%CI,85.3-97.2%),respectively.5.The results of Delong test showed that the discrimination performance of the nomogram was better than the CF and RF models in the two cohorts.However,the difference between the nomogram and PP_Radscoreoringinal models was no statistical significance.6.The analysis results of the calibration curve and DCA demonstrated that the nomogram had an excellent goodness of fit and the highest clinical application for MVI prediction.Conclusion:The radiomics features are highly objective and automated.The CECT-based tumor radscores can be used as potential image-biomarkers for preoperative MVI prediction.A nomogram based on radiomics-clinical-radiological features can be used to assess the status of MVI in a non-invasive,highly accurate and repeatable way,which may provide a new pattern for personalized preoperative risk assessment and clinical management of HCC patients.