Radiomics Models Based On Enhanced CT In Predicting Histological Subtypes Of Thymomas

Objective To explore the value of radiomics model based on enhanced CT in predicting histological subtypes of thymomas.Materials and Methods 226 patients with thymomas confirmed by pathological diagnosis were analyzed retrospectively,there were 146 cases low-risk group thymomas,(type A,type AB,type B1 are 23 cases,96 cases,27 cases),and 80 cases high-risk group thymomas,(type B2,type B3 are 53 cases,27 cases),they were divided into training dataset(159 cases)and test dataset(67cases)by a ratio of 7 to 3.ITK-SNAP software was used to manually delineate the region of interest(ROI)on the arterial phase(AP)and venous phase(VP)images of CT,extract radiomics features by Pyradiomics software and all cases were standardized by Image Biomarker Standardization Initiative(IBSI),then used minimum redundancy maximum relevance(m RMR)and the least absolute shrinkage and selection operator(LASSO)selected radiomics features of AP,VP related to thymomas histological subtypes to build AP radiomics model and VP radiomics model,then combine the AP and VP and reselect the radiomics features to build arterial-venous phase(A-VP)radiomics model,Receiver operating characteristic(ROC)curve and Area under the curve(AUC)to evaluate the prediction efficiency of three models,and then split the whole data set randomly 100 times,different training and test datasets were established to compare the prediction efficiency of three models(AUC value,Accuracy,Sensitivity,Specificity).Compare general clinical and CT features of different histological subtypes of thymomas,there was significant difference when P<0.05,independent clinical and imaging features related to histological subtypes were identified by logistic regression analysis and a clinical model was built,the clinical features include gender,age and Myasthenia Gravis(MG),the CT features include Dmax,location,morphology,boundary,cystic necrosis,calcification,invasion,metastasis,the CT value of AP and VP,enhancement uniformity and enhancement mode.Combine radiomics features with clinical and CT features,multivariate logistic regression analysis was used to establish the radiomics nomogram,ROC and AUC value to evaluate the prediction efficiency of radiomics nomogram,De Long’s test to compare the AUC values of radiomics model,clinical model and radiomics nomogram.Hosmer–Lemeshow test and Decision curve analyze(DCA)to evaluate and validate the nomogram results.Results 788 features were extracted from the AP and the VP respectively,3 radiomics features were selected to build AP radiomics model and the AUC value of this model in training dataset and test dataset were 0.78(95%CI: 0.70-0.85),0.76(95%CI: 0.65-0.88);17radiomics features were selected to build VP radiomics model and the AUC value of this model in training dataset and test dataset were 0.85(95%CI: 0.79-0.92),0.84(95%CI: 0.74-0.93);19 radiomics features were selected to build A-VP radiomics model and the AUC value of this model in training dataset and test dataset were 0.89(95%CI: 0.84-0.94),0.85(95%CI: 0.75-0.94);in training dataset,A-VP radiomics model had a better performance than AP radiomics model and VP radiomics model in AUC value,Accuracy,Sensitivity and Specificity(P<0.05),VP radiomics model had a better performance than AP radiomics model in AUC value,Accuracy,Sensitivity and Specificity(P<0.05),in test dataset,there was no significant difference between three radiomics models in accuracy,sensitivity and specificity,VP radiomics model and A-VP radiomics model had a better performance than AP radiomics model in AUC value(P<0.05),there was no significant difference between VP radiomics model and A-VP phase radiomics model in AUC value(P>0.05).There was significant difference between low-risk group thymomas and highrisk group thymomas in gender,age and MG,morphology,boundary,calcification,invasion,metastasis and the CT value of AP and VP(P<0.05);there was no significant difference in Dmax,location,cystic necrosis,enhancement uniformity and enhancement mode(P>0.05),the AUC value of clinical model in training dataset and test dataset were 0.79(95%CI: 0.71-0.86),0.81(95%CI: 0.70-0.91).The AUC value of radiomics nomogram in training dataset and test dataset were 0.91(95%CI: 0.86-0.95),0.88(95%CI: 0.80-0.96),in training dataset,radiomics nomogram and radiomics model had a better performance than clinical model in AUC value(P<0.05),there was no significant difference between radiomics nomogram and radiomics model in AUC value(P>0.05),in test dataset,there was no significant difference between radiomics nomogram,radiomics model and clinical model in AUC value(P>0.05).DCA shows that taken 0.1-1 as the threshold of probability,the clinical usefulness of radiomics nomogram had a better performance than radiomics model and clinical model.Conclusion(1)Radiomics models based on Enhanced CT had a good performance in predicting thymomas histological subtypes.(2)VP radiomics model and A-VP radiomics model had a better performance than AP radiomics model.(3)The clinical usefulness of radiomics nomogram had a better performance than radiomics model and clinical model.