关键词: 胶质母细胞瘤, 原发性中枢神经系统淋巴瘤, 多模态MRI, 影像组学

Abstract: Objective To explore the value of multimodal MRI radiomics technology in accurately distinguishing glioblastoma (GBM) from primary central nervous system lymphoma (PCNSL). Methods Patients diagnosed with GBM (n=52) or PCNSL(n=33) through surgery at our hospital from May 2021 to June 2024 were selected as the research subjects.Radiomics features were extracted from contrast enhancement-T1 weighted imaging (CE-T1),T2WI,and T2-fluid-attenuated inversion-recovery (T2-Flair) sequences,respectively.Features were screened through F-test,Pearson correlation coefficient,and LASSO.K-nearest neighbor (KNN),random forest (RF),support vector machine (SVM),logistic regression (LR),linear discriminant analysis radiomics features,and 5-fold cross validation (5-CV) was performed.Receiver operating characteristic (ROC) curve,calibration curve,and decision curve analysis (DCA) curve were used to evaluate the effectiveness of the model. Results There were statistically differences in tumor location,number of lesions,recurrence/primary,and MRI characteristics (obvious uniform enhancement,cystic changes,necrosis,and bleeding) between GBM and PCNSL patients (P＜0.05).After screening,CE-T1,T2WI,and T2-Flair sequences retained 15,20,and 15 features,respectively,while multiparametric MRI retained 26 features.The average accuracy of the KNN classifier multi parameter MRI model was the highest;The SVM classifier had the highest average recall rate for the multi parameter MRI model;The neural network classifier multi parameter MRI model had the highest average accuracy and F1 score.ROC curve analysis showed that the LDA classifier CE-T1 (0.916) and T2WI (0.930) models had the highest AUC;The SVM classifier T2-Flair model (0.865) has the highest AUC;The LR classifier had the highest AUC for the multi parameter MRI model (0.956);Among the six classifiers,the AUC of the multi parameter MRI model was greater than that of the single sequence model.DeLong test showed that the AUC of multi parameter MRI models differed significantly from that of any single sequence model in KNN,RF,LR,and neural networks (P＜0.05).Calibration curve showed that the RF classifier's multi parameter MRI model's predictions matched the actual probabilities well.DCA analysis showed that all six classifier multi parameter MRI models had good clinical benefits. Conclusion The prediction model constructed by combining multi parameter MRI with machine learning (ML) classifier can accurately distinguish GBM from PCNSL.The discrimination,prediction accuracy,and clinical value of multi parameter MRI models in each classifier are superior to those of single sequence models.

Key words: Glioblastoma, Primary central nervous system lymphoma, Multimodal MRI, Radiomics