Machine learning-based mri radiomics for predicting bevacizumab response in peritumoral brain edema of glioblastoma

To explore the prediction of the efficacy of bevacizumab (BEV) in treating peritumoral brain edema based on the radiomic features of T1-enhanced and T2-FLAIR MRI in patients with glioblastoma.

A retrospective study was conducted on 321 glioblastoma patients who received bevacizumab treatment at The First Affiliated Hospital of Soochow University and The Second Affiliated Hospital of Soochow University from January 2020 to January 2025. The patients were randomly divided into a training set (n = 224) and a validation set (n = 97) at a ratio of 7:3. According to the Response Assessment in Neuro-Oncology (RANO) criteria, patients were classified into the remission group and non-remission group based on the reduction rate of peritumoral edema volume. Radiomic features were extracted from the glioma enhanced tumor region and peritumoral edema region on pretreatment baseline MRI using 3D-Slicer software. Features with significant intergroup differences were screened by univariate analysis and Lasso regression. Subsequently, three machine learning algorithms, namely logistic regression (LR), support vector machine (SVM), and random forest (RF), were used to construct radiomics prediction models. The efficacy of each model was compared using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). The clinical practicability of the models was further evaluated by decision curve analysis.

A total of 9 optimal radiomic features derived from the enhanced tumor and peritumoral edema regions were selected to build the machine learning models. In the training set, the AUC values of the LR, SVM, and RF models were 0.832 (95% confidence interval [CI]: 0.690-0.973), 0.814 (95%CI: 0.730-0.897), and 0.796 (95%CI: 0.707-0.883), respectively. In the validation set, the corresponding AUC values were 0.883 (95%CI: 0.769-0.996), 0.800 (95%CI: 0.713-0.886), and 0.770 (95%CI: 0.604-0.934), respectively. For the LR model, decision curve analysis showed that the net benefit of the model was greater than 0 when the risk threshold ranged from 0.05 to 0.95.

The LR, SVM, and RF models built on MRI radiomic features from the combined enhancing tumor and peritumoral edema region all showed promising predictive performance. Among them, the LR model demonstrated favorable clinical net benefit in decision curve analysis, suggesting potential clinical utility. These results indicate that the LR model may help predict bevacizumab response for peritumoral edema in patients with glioblastoma.