The Application Of CT-based Radiomics In Clinical Evaluation Of Hepatic Alveolar Echinococcosis

Objective:To delineate the region of interest for hepatic alveolar echinococcosis（AE）lesions in CT image,extract and screen the best imaging features.Furtherly,to establish the radiomics prediction models（biological activity assessment,extrahepatic invasion and metastasis prediction,microvessel density status evaluation）and evaluate the prediction efficiency of these models.Methods:174 patients of hepatic AE were included retrospectively.According to the evaluation of biological activity by PET/CT,the patients were divided into active group and inactive group.The regions of interest of hepatic AE lesions on CT images were extracted and screened,and the prediction model was established.The prediction efficiency of the model was evaluated by receiver operating characteristic curve（ROC）and the Delong test was used to screen the best model Mean while,the clinical indexes related to the biological activity were screened,the clinical-radiomics nomogram was established,and the model was evaluated by calibration curve and decision curve analysis.Delong test was used to evaluate whether there was a difference in the prediction efficiency between radiomics model and combined model.The data of 201 hepatic AE patients were collected at the same time.The patients were grouped according to whether there were extrahepatic invasion and/or metastasis.The logistic regression classifier was used to establish the prediction model,ROC curve was used to evaluate the prediction efficiency of the model,the calibration curve and decision curve were used to evaluate the clinical practicability.The clinical and imaging data of103 hepatic AE patients who underwent surgical resection were collected.The patients were grouped according to pathologic microvessel density（≤15 and>15）in the marginal area of AE lesions and then the radiomics models were established.The prediction efficiency of the models was evaluated by ROC curve.The clinical features were screened.Univariate and multivariate logistic regression analysis were performed to screen the independent risk factors for predicting the microvessel density status in the marginal area of hepatic AE.Result:In the prediction of hepatic AE biological activity,three classifiers—k nearest neighbor,logistic regression and multilayer perceptron were selected to establish the prediction model.In the training set,the area under curves（AUC）of ROC curves were 0.827,0.844 and 0.952 respectively,and the test set were 0.748,0.796 and 0.800 respectively;Delong test showed that the prediction model established by multilayer perceptron was the best model(P_{KNN-MLP-train}=0.0001,P_KNN-MLP-test=0.0355,P_LR-MLP-train<0.0001,P_LR-MLP-test=0.8941).In the screening of clinical indicators,univariate analysis of the training set showed that there were significant differences in P stage,small vesicle sign and calcification between the active and inactive groups（P<0.05）.The AUC of Rad-score combined clinical model for predicting the biological activity in the training set and the test set were 0.816（P<0.0001）and 0.796（P=0.0100）,respectively.Delong test showed that there was no significant difference in the prediction effect between the combined model and the single radiomics model(P_train=0.8764,P_test=0.0623).In the prediction of extrahepatic invasion and metastasis,the AUC of the logistic regression model in the training set was 0.996,the accuracy was 0.989,the sensitivity is0.989 and the specificity was 0.989;in the test set,AUC was 0.705,accuracy was 0.705,sensitivity was 0.795 and specificity was 0.615.The calibration curve showed that the radiomics model had overestimated risk,that is,some patients without extrahepatic AE lesions were judged to extrahepatic group.The decision curve showed that the radiomics model has good practicability in the training set.When the risk threshold probability of patients in the test set was in the range of 0.23-0.67,the model could still increase the benefit of the AE patients.In the prediction for microvessel density status,four classifiers（k nearest neighbor,logistic regression,multi-layer perceptron and support vector machine）were selected to establish the prediction model.In the training set,the AUCs were 0.816,0.761,0.778 and 0.748 respectively,and in the test set,the AUCs were 0.741,0.750,0.741 and 0.806 respectively;the accuracy of training set were 0.744,0.744,0.720 and0.732 respectively,and the accuracy of in test set were 0.762,0.810,0.810 and 0.810respectively.Univariate analysis showed that there were significant differences in the detection of Eg B antibody,Eg CF antibody and the maximum diameter between the two groups of microvessel density（P<0.05）.Univariate logistic regression analysis showed that the maximum diameter（P=0.003）and rad-score（P=0.0015）were positively correlated with microvessel density.Multivariate logistic regression showed that only rad-score（P=0.0175）was positively correlated with microvessel density.Conclusion:1)The prediction model of hepatic AE biological activity based on CT image has good prediction efficiency,but the combined clinical-radiomics model can not significantly improve the prediction efficiency;2)In the prediction model of extrahepatic invasion and metastasis of AE,the logistic regression model has good prediction efficiency and has good clinical utility,but it overestimates the risk of extrahepatic lesions;3)In the prediction of microvessel density status of AE lesions,the prediction models established by the four classifiers have good prediction effect,and the support vector machine model performs the best in test set.