Abstract: Objective To explore the efficacy of different radiomic models based on susceptibility-weighted imaging(SWI)sequences in predicting the isocitrate dehydrogenase(IDH)mutation status of glioma before operation. Methods A retrospective analysis was conducted on the imaging data of 493 adult patients with confirmed diffuse glioma from UCSF-PDGM, including 393 wild-type cases and 100 mutation cases. The patients were divided into training(n=395)and testing(n=98)sets in an 8∶2 ratio. Radiomic features were extracted from SWI sequences based on two regions of interests(ROIs): tumor core(TC)and whole tumor(WT). A total of 1 316 radiomic features were standardized using the z-score method, and dimensionality reduction was performed using principal component analysis(PCA). Feature selection was conducted via variance analysis. Support vector machine(SVM), linear discriminant analysis(LDA), auto-encoder(AE), Logistic regression(LR), Logistic regression via Lasso(LR-Lasso), and native bayes(NB)models were constructed. Receiver operating characteristic(ROC)curves were drawn to assess the accuracy, sensitivity, and specificity. The models performance was evaluated using the area under the curve(AUC)on the testing set. Results A total of 20 radiomic features were selected to establish the radiomic model. The AUC and accuracy of the SVM model were 0.841 and 0.755; the AUC and accuracy of the LDA model were 0.800 and 0.735; the AUC and accuracy of the AE model were 0.743 and 0.745; the AUC and accuracy of the LR model were 0.842 and 0.725; the AUC and accuracy of the LR-LASSO model were 0.880 and 0.857; the AUC and accuracy of the NB model were 0.806 and 0.725. Conclusion SWI-based radiomic features hold certain value in predicting IDH gene mutations in glioma. The LR-Lasso model demonstrates the best predictive performance among the models.

Key words: Susceptibility-weighted imaging, Isocitrate dehydrogenase, Glioma, Radiomics, Machine learning