Purpose: Recently, quantitative analysis based on magnetic resonance images (MRI) has played an important role in glioma grading that has clinical significance in treatment strategy and prognosis prediction. Therefore, this study aimed to find out discriminative three-dimensional (3D) textural features extracted from multimodal MRI for grading glioma. Furthermore, an automatic glioma grading strategy using the discriminating features was proposed to provide valuable diagnostic information to clinical practice. Materials and Methods: In this study, 143 patients with postoperatively pathologically confirmed glioma were enrolled (39 and 104 low- and high-grade, respectively). For each patient, volumes of interest (VOIs) were delineated on multimodal MRI (structure, diffusion and perfusion images). 3D gray-level co-occurrence and curvature co-occurrence matrix (GLCM and GLGCM) textural features extracted from the VOIs were innovatively used to describe the intra-glioma heterogeneity in this study. Then a support vector machine (SVM) based sample augmentation, feature selection and classification strategies were proposed to firstly obtain an optimal feature set and then verify its capacity to grade glioma, comparing with histogram features often used in previous studies. The McNemar test (p<0.05) and Kappa score were implemented to compare the grading performance of each feature groups with the pathological grading results. Results: For each patient, a total of 410 features were generated. By the adapted SVM-recursive feature elimination method, 59 optimal features were determined, which includes 20, 22 and 17 GLCM and GLGCM features from structure, diffusion and perfusion modalities, respectively. Using the SVM classifier, this optimal feature set achieved the best classification performance in discriminate low- and high-grade glioma, comparing with histogram features. According to the McNemar test result and Kappa score, the grading results of optimal feature set were strongly consistent with pathological results. The area under receiver operating characteristic curve, accuracy, sensitivity and specificity were 0.9884, 95.67%, 96.15% and 95.19%, respectively. Conclusion: 3D GLCM and GLGCM textural features from VOIs of multimodal MRI were outstanding image biomarkers in identify subjects with different glioma grades. Features from different types of modalities indicated that different structural, diffusion and perfusion pattern in 3D glioma lesions all contributed to the grading of glioma. This automatic glioma grading strategy proposed in our study is feasible and valuable in clinical practice of preoperative glioma diagnosis.