| Objective: To explore the performance of radiomics models based on multi-sequence MRI in predicting histological grade,Ki67 expression and myometrial invasion of endometrial carcinoma(EC).Materials and Methods: The data of 306 patients with endometrial carcinoma(EC),enrolled in Tianjin Medical University Cancer Hospital from December 2013 to January 2019,confirmed by surgical pathology were analyzed retrospectively in this study.All patients underwent preoperative pelvic MR scan and contrast-enhanced MR.Two radiologists used ITK-SNAP software to delineate and segment region of interest(ROI)layer by layer in pelvic MR imaging,which included T2 WI,DWI,apparent diffusion coefficient(ADC)images and DCE-MRI delayed images and all images are sagittal images.(1)Radiomics features were extracted by using radiomics software based on Matlab2016 a and 459 features are extracted from each sequence.Pearson correlation coefficient and Kruscal-wallis rank sum test were used for feature selecting.The cases are randomly divided into training set and validation set according different tasks.The selected features are incorporated into the elastic network model and the super-parameter α value of the model with the best efficiency is obtained through cross-verification,and the model is initialized.The best super-parameter λ is determined in the training set through the ten-fold cross-validation method to construct the radiomics model.The receiver operating characteristic(ROC)curve is used to evaluate the prediction ability of the radiomics model,and the validation set data are used to verify the model.(2)The linear equation is constructed using the coefficients of each selected feature under the best super-parameter λ and the radscore is calculated.The radscore of two radiomics models were used to predict myometrial infiltration,and its prediction performance was evaluated by ROC curve.Results: Among the 306 EC patients,there were 232 cases of high grade(G2,G3),74 cases of low grade(G1),137 cases of Ki67 high proliferation group(≥50%)and 121 cases of low proliferation group(<50%).There were 57 cases of deep myometrial infiltration.Training sets and validation sets were randomly divided according to different training tasks,and there was no significant difference in basic information between training sets and validation sets.2.(1)A total of 33 radiomics features were finally incorporated into the elastic network regression model through feature selecting.(2)Through 10-fold cross-validation,it is calculated that when α is zero,all of the selected features are incorporated in the ridge regression,the model has the best efficiency.(3)The best hyperparameter λ,which represents L2 regular pattern,was obtained through 10-fold cross-validation to construct the best prediction model.The area under ROC curve(AUC)of the histological grading prediction model reached 0.769 with a sensitivity of 84.5% and a specificity of 61.8% in training set,and the AUC in validation set reached 0.678 with a sensitivity of 58.3% and a specificity of 78.4%.The AUC of the Ki67 expression prediction model reached 0.714 with a sensitivity of 75.9% and a specificity of 56.7% in training set,and the AUC in validation set reached 0.733 with a sensitivity of 79.2% and a specificity of 71.4%.3.The radscores of the two radiomics models were both significantly different in different myometrial infiltration groups(all P < 0.001).The AUC of evaluating the myometrial invasion by the histological grading prediction model reached 0.710 with a sensitivity of 78.9% and a specificity of 56.2%,and the AUC of evaluating by the Ki67 expression prediction model reached 0.691 with a sensitivity of 93.3% and a specificity of 42.7%.Conclusion: The radiomics model based on multi-sequence MRI has important value in preoperationally predicting histological grade,Ki67 expression and myometrial invasion of EC,and it is helpful to assist clinicians to realize accurate and personalized treatment as soon as possible. |
PDF Full Text Request