CT-based Radiomics For Preoperative Prediction Of Microvascular Invasion In Hepatocellular Carcinoma

Backgroundhepatocellular carcinoma(HCC)is the sixth most common cancer and the third most common cause of cancer death in the world.Microvascular invasion(MVI)is a risk factor for early recurrence and poor prognosis of HCC.Early prediction of liver cancer MVI can reduce the recurrence rate,but MVI can only be diagnosed by postoperative histopathology,so its clinical practicability is limited.Imageomics is the latest popular imaging analysis technology,which aims to find and translate the uncoded information in medical images.By constructing an appropriate model with fine characteristics,it has obtained good evaluation and prediction ability in various difficult clinical tasks.At present,many studies have applied imaging methods to the diagnosis of MVI of liver cancer,which will help to formulate flxrther treatment plans and provide early intervention measures for patients,and effectively reduce the postoperative recurrence rate of liver cancer.ObjectiveTo predict MVI of liver cancer based on image feature model.MVI of liver cancer was predicted based on CT imaging and nomogram model.MethodsThe patients involved in the study were recruited in Nanfang Hospital(institution I)of Guangzhou Southern Medical University and Jianggao District of Guangzhou southern hospital(institution II).We retrospectively searched all HCC cases diagnosed by postoperative pathology from January 2017 to June 2021,and a total of 768 HCC patients were adopted in the electronic medical record.318 cases were identified after screening by inclusion and exclusion criteria.Among them,patients from institution I were randomly divided into training group(n = 236)and validation group(n = 55)to construct and test the proposed classifier.The data set from agency II was used as an independent test queue(n = 27).We collected four phase CT data of patients with hepatocellular carcinoma,including plain scan phase,arterial phase,portal phase and delayed phase.Three tumor targets VOItumor,VOItum〇r+icm and VOItum〇r+2cm were generated by semi-automatic segmentation,Image radiomics features were extracted.PCA principal component analysis and logistic regression algorithm are used to reduce the dimension,feature selection and construct radiomics labels of the training set.Univariate and multivariate analyses were used to identify important clinical factors and imaging features associated with MVI and then included in the predictive nomogram.The performance of imaging histograms is evaluated by their calibration,identification and clinical application.ResultsAFP,non smooth margin,capsule defect and internal artery were independent predictors of MVI.The AUC of the single imaging feature model was 0.687(95%Cl 0.617-0.757)in the training cohort,0.659(95% Cl 0.588-0.731)in the validation cohort and 0.626(95% Cl 0.553-0.699)in the test cohort.The AUC of the model constructed by clinical factors and imaging features was 0.766(95%CI0.703-0.827)in the training group,0.759(95%CI 0.619-0.878)in the validation group and 0.712(95%CI 0.486-0.908)in the test group.The prediction ability of VOItum〇r+2cm radiomics in training cohort,validation cohort and test cohort is better than VOItumor and VOItum〇r+icm(training cohort: 0.806(95%CI 0.748-0.859)5verification cohort: 0.839(95%CI 0.707-0.938)and test cohort: 0.800(95%CI0.600-0.955)).The AUC of the prediction model including,VOItum〇r+2cm radiomics features,clinical features and imaging features in the training group was 0.859(95%CI 0.810-0.903),the validation group was 0.847(95%CI 0.728-0.940)and the test group was 0.841(95%CI 0.656-1.000).ConclusionOur study used imaging techniques to predict the occurrence of MVI.VOItum〇r+2cm radiomics model has stronger prediction ability than other radiomics models.The nomogram of nomogram model(clinical features,imaging features and radiomics features)is integrated to satisfactorily predict the individualized risk of MVI in HCC patients.