Preliminary Study HAE Activity Based On CT Radiomics Featuresevaluation

Objective: To explore the value of CTradiomics features in evaluating the biological activity of HAE lesions.Methods: Retrospective analysis was performed on 163 HAE patients,including 70 males and 93 females,with an average age of 42±12 years,who were confirmed by surgerical pathology and received abdominal CT enhanced scan and PET-CT examination before surgery.Among them,107 patients had biologically active lesions and 56 patients had no biologically active lesions(SUVmax ≥5 was defined as biologically active).Based on radiomics,the original CT images in patients are imported into Rad Cloud platform,to draw lesions in portalvenousphase(PVP)of the image,getvolumeofinterest(VOI)and extract the radiomics features,then using dimension-reduction methods such as LASSO algorithm(least absolute shrinkage and selection operator)to choose the valuable characteristics of biological activities.Smote algorithm is used to increase the sample size of a few classes(no biological activity),Three classifier learning models include Logical regression(LR),multilayer perceptron(MLP)and random forest(RF)were established to predict the biological activity of lesions.The training data set was used for training,and the test data set was used to test the model performance.The performance of the classifier was mainly evaluated by the area under the receiver operating characteristic curve(ROC)curve(AUC)and accuracy.Results: 1409 radiomics features were extracted,13 biological-related feature parameters were obtained by dimensionality reduction.The RF machine learning model was used to obtain the best performance among the three methods,and the test set(AUC[area under curve],0.817;95% confidence interval [CI],0.685-0.910;The accuracy rate was 80.77%),showing good classification performance.Conclusion: It is of great value to evaluate the biological activity of HAE lesions based on CT radiomics features.