Prediction Of Microvascular Invasion In Hepatocellular Carcinoma Based On MRI Texture Analysis

ObjectiveThis research seeks to evaluate the efficacy of Magnetic Resonance Imaging(MRI)radiomics in forecasting microvascular invasion(MVI)in hepatocellular carcinoma(HCC),and to supply a benchmark for the selection of clinical treatments and the assessment of prognosis.MethodsRetrospective analysis of 77 patients with postoperative pathological findings of hepatocellular carcinoma(HCC)surgically resected in Affiliated Hospital of Chifeng University from 07/2018 to 05/2022.The patients were divided into two groups according to the postoperative pathological findings: 26 patients with MVI(+)and 51 patients with MVI(-).One month before surgery,all patients underwent plain and dynamic contrast-enhanced MRI(DCE-MRI)examinations at our hospital.The imaging characteristics of the tumor were measured and evaluated on the patient’s T2WI,DWI,and DCE-MRI(mask,late arterial phase,portal phase,and delay phase)images,including the maximum diameter of the tumor,whether the tumor margin was regular,whether the envelope was intact,and whether abnormal enhancement around the tumor,etc.General information such as the patient’s age and gender,as well as the tumor marker alpha-fetoprotein(AFP)value.Independent samples t-test and independent samples χ~2 test were used to statistically analyze the characteristics such as gender,age,AFP value,maximum tumor diameter,whether the tumor margin was regular,whether the envelope was intact,abnormal enhancement around the tumor,and to select the characteristics with statistically significant differences between the MVI(+)and MVI(-)groups,and to establish a prediction model(clinical model)using multi-factor logistic regression.The regions of interest(ROI)of HCC lesions on T2WI,DWI,masked,late arterial,portal,and delayed MRI image images were manually outlined using 3D-Slicer(Slicer 4.11.20210226)software,and the imaging histological features were extracted.By employing the least absolute shrinkage and selection operator(LASSO),dimensionality was reduced and the most pertinent imaging histological features were filtered out with MVI,and a prediction model was then constructed through logistic regression(Binary Logistic),including a single MRI sequence-based.The imaging histology model based on a single MRI sequence and the imaging histology model based on the above six sequences were used.The imaging histology model including the above-mentioned 6 sequences was established by fusing the information of each single sequence texture feature.Finally,the clinical features and imaging histology features were combined to build prediction models(fusion models)using logistic regression(Binary Logistic).The prediction model equations were used to construct the receiver operating characteristic curve(ROC),and area under the curve(AUC)was determined for assessment.The effectiveness of the model in predicting the MVI state in HCC was evaluated.The validation method used Bootstrap method(enhancement method iterative boot1000)to internally validate the MVI status of HCC in the group of patients.Results1、Clinical data,imaging features and clinical modelThe differences in clinical data AFP values and imaging features in the maximum tumor diameter,whether the lesion margin was regular,whether the tumor envelope was intact,and whether there was abnormal enhancement around the tumor were statistically different between the MVI(+)group and the MVI(-)group(p<0.05).The efficacy of the clinical model based on the above characteristics to predict the MVI status of HCC: training set AUC=0.770(95%CI: 0.701,0.843),sensitivity 0.813,specificity 0.935,accuracy 0.748.Bootstrap internal validation AUC=0.763(95% CI: 0.708,0.821).2.Imaging histology modelThe T2WI,DWI,mask,arterial phase,portal phase,and delayed phase could be screened with 2-6 texture features,respectively,and the corresponding prediction models were established based on the respective texture features,and the efficacy of predicting MVI status in HCC was 0.773(95% CI: 0.710,0.845)for T2WI,0.853 for sensitivity,0.865 for specificity,0.718 for accuracy,internal validation 0.768(95% CI: 0.702,0.843);DWI 0.760(95%CI: 0.691,0.838),sensitivity 0.821,specificity 0.883,accuracy 0.704,internal validation: 0.751(95% CI: 0.676,0.833);montages 0.752(95% CI: 0.661,0.840),sensitivity 0.854,specificity 0.862,accuracy 0.716,internal validation0.736(95% CI: 0.664,0.832);arterial phase 0.788(95% CI: 0.732,0.847),sensitivity 0.885,specificity 0.824,accuracy 0.709,internal validation 0.778(95% CI: 0.724,0.829);portal phase 0.813(95% CI: 0.755,0.849),sensitivity0.837,specificity 0.890,accuracy 0.727,internal validation 0.804(95% CI:0.745,0.829);delayed phase 0.785(95% CI: 0.733,0.841),sensitivity 0.852,specificity 0.882,accuracy 0.734,internal validation 0.757(95% CI: 0.722,0.816).In total,eight imaging histological features were screened and the imaging histological model was established: predictive efficacy 0.829(95% CI: 0.793,0.870),sensitivity 0.874,specificity 0.880,accuracy 0.754,internal validation0.805(95% CI: 0.762,0.846).3.Fusion modelThe fusion model predicted MVI efficacy 0.853(95% CI: 0.811,0.874),sensitivity 0.930,specificity 0.881,accuracy 0.811,internal validation 0.824(95% CI: 0.806,0.859)for HCC.ConclusionPreoperative AFP,the maximum diameter of the tumor,the integrity of the pseudocapsule,the regularity of the tumor edge,and the abnormal enhancement around the tumor are all independent risk factors for predicting MVI in HCC.The clinical model,radiomics model and fusion model all have certain efficacy in predicting MVI status in HCC,and the fusion model has the highest efficacy.