Preoperative Prediction Of Microvascular Invasion Risk Grades In Hepatocellular Carcinoma Based On Tumor And Peritumor Dual-region Radiomics Signatures

ObjectivesTo explore the impact of dual-region radiomics signatures of the tumor and peritumor on microvascular invasion(MVI)prediction,this study aimed to develop enhanced MRI radiomics models with different regions of interest(ROIs).Material and methodsA total of 501 patients with hepatocellular carcinoma who underwent preoperative Gd-EOB-DTPA-enhanced MRI and radical hepatectomy within 1 month(training set n=402,test set= 99)were analyzed retrospectively.The tumor lesions were labeld on T1-weighted imaging(T1WI),arterial phase(AP),portal venous phase(PVP)and hepatobiliary phase(HBP)images,and then the original tumor lesions were expanded by 10 mm and 20 mm in u Al-Research-Portal(Shanghai Lianying Intelligent Medical Technology Co.,Ltd.).The labeled and expanded images were imported into u Al-Research-Portal.First,the image was preprocessed,and then the radiomics features within the labeled range of the lesions were extracted.The least absolute shrinkage and selection operator(Lasso)was used to select the tumor and peri-tumor features that were important for MVI risk classification,and then Berkes-Cox transformation was performed.Based on tumor,tumor and peri-tumor 10 mm,tumor and peri-tumor 20 mm,a single modal radiomics model was established by Logistic algorithm on the training set samples,and its performance of predicting MVI was tested on the test set.Using the method of combinatorial modeling,the best multimodal radiomics model of ROI based on tumor is selected according to AUC,and then the multimodal radiomics models of ROIs based on tumor and peri-tumor 10 mm,tumor and peritumoral20 mm is established accordingly.The Lasso algorithm is used to select the clinical and radiological features which are important to MVI.Combined with the radiomics features,a single-modal clinico-radiological radiomics model is established,and the above combined modeling method is used to establish a multimodal clinico-radiological radiomics model.Finally,the AUC values of the above prediction models were compared,and the best prediction model of microvascular invasion risk classification of hepatocellular carcinoma was selected.According to the prediction results of the best model,the Kaplan-Meier method was used to draw the survival curve of HCC patients with different MVI grades.Results1.Clinical and radiological features were selected by Lasso algorithm,and 15 important clinical and radiological features were selected for MVI classification,including serum AFP level,Child-Pugh,liver cirrhosis,age,prothrombin time,platelet count,shape,hepatobiliary phase low signal intensity,intratumoral hemorrhage,satellite focus,diameter,number of nodules,arterial phase peritumoral enhancement,capsule,tumor diameter.2.In the test set,the AUC values of dual-regions(tumor and peri-tumor 20mm)based on AP,PVP,T1 WI and HBP images(AP(20),PVP(20),T1WI(20),HBP(20))and single-region(tumor)models(AP(0),PVP(0),T1WI(0),HBP(0))were 0.741 vs0.694,0.733 vs0.725,0.667 vs 0.710,and 0.559 vs 0.677,respectively.3.Among multimodal radiomics models based on ROI of tumor,the radiomics model(T1WI(0)& PVP(0)& AP(0))performed best,and the AUC value of the test set was 0.758 and 0.616.In the corresponding dual-region(tumor & peritumoral 10 mm or tumor&peritumoral 20mm)radiomics model created by combinatorial modeling,Multimodal radiomics model(T1WI(20)& PVP(20)& AP(20))based on tumor and peritumoral 20 mm showed better ability in predicting MVI risk classification,and the test set AUC value was0.778 and 0.636.4.The multimodal prediction model(T1WI(20)& AP(20)& PVP(20))with clinical radiation features is more effective than the corresponding multimodal radiomics prediction model.In the test set,AUC: 0.852 vs 0.778ACC: 0.747 vs 0.737.5.The recurrence-free survival time(RFS)predicted by the final model in M0,M1 and M2 groups was significantly different,and the median recurrence-free survival time was 37 months,27 months and 8 months respectively(p < 0.001).Conclusion1.In predicting the risk classification of microvascular invasion,peritumoral radiomics features on arterial phase and portal venous phase can provide supplementary information for tumor radiomics features.2.The inclusion of the radiomics characteristics of peri-tumor larger than 1cm is meaningful for the prediction of MVI grade.3.The predictive model established in this study can effectively predict MVI risk classification and stratify the prognosis of patients with HCC.