RESUMEN
Objective To observe the value of preoperative MRI radiomics models for predicting risk stratification of endometrial cancer(EC).Methods Data of 219 EC patients who underwent pelvic MR examination before surgery were retrospectively analyzed.The patients were divided into high risk group(n=104)or low risk group(n=115)according to postoperative pathological findings,also assigned into training set(n=153)or test set(n=66)according to examination time and further divided into high or low risk subgroups in each set.ROI was manually sketched on MRI using 3D Slicer,and each 1 130 features were extracted from axial and sagittal fat suppression(FS)T2WI as well as axial and sagittal enhanced FS-T1WI,respectively.Then the least absolute shrinkage and selection operator(LASSO)was used to select a total of 54 merged MRI features,including 12,14,16 and 12 features,respectively.Finally,25 merged LASSO features were reduced dimensionality and selected by reusing LASSO.Extremely randomized trees algorithm was used to construct radiomics models based on each single sequence features,merged MRI features and merged LASSO features,respectively.Receiver operating characteristic curves were drawn,the area under the curve(AUC),the accuracy and F1 score were obtained to evaluate the predicting efficacy of each model.AUC was used to evaluate the predictive efficacy of the models and subjective diagnosis of test set.Results In training set,the accuracy(0.784,0.777),F1 score(0.730,0.731)and AUC(0.835,0.855)of modelmerged MRI and modelmerged LASSO were both higher than those of each single sequence model,while in test set,the sensitivity(0.794,0.882),specificity(0.909,0.969)and AUC(0.904,0.934)of modelmerged MRI and modelmerged LASSO were both higher than those of subjective diagnosis and each single sequence model.The predictive effiency of modelmerged LAsSo was better than that of modelmerged MRI,which was the best model.Conclusion Preoperative MRI radiomics model was effective for predicting risk stratification of endometrial cancer.Modelmerged LASSO had the best performance.