Purpose: The proposed study aims to develop an MRI-based radiomics analysis framework and investigate the feasibility of the calculated quantitative imaging features for differentiating low from high grade soft tissue sarcomas (STSs). Material and Methods: A total of 22 patients (9 low grade and 13 high grade) who were pathologically diagnosed with soft tissue sarcomas were recruited for the analysis and corresponding T2-weighted MR images were acquired for further post-processing. Tumour delineations were manually traced slice by slice concluding to whole tumour annotated volumes from all enrolled patients. A total of 1165 high-throughput patient-specific quantitative imaging features were exported from each volume using radiomics and evaluated using random forest machine learning classifiers. The overall analysis framework was coupled with feature selection and oversampling techniques to address high-dimensionality dataset issues and the unbalanced ratio between the two examined groups. Validation was performed using repeated nested cross-validation to eliminate overfitting problems and assess the stability of the classification performance. Results: The classifier, using the three most important radiomic features selected though training, yielded an accuracy of 0.781 ± 0.15, an area under the receiver operating characteristic curve (AUROC) equal to 0.814 ± 0.186, F1-score of 0.704 ± 0.198, 0.762 ± 0.267 and 0.725 ± 0.283 for precision and recall respectively using multiple independent test sets. Conclusions: Radiomic features from routine MR imaging protocols can provide a strong discriminatory performance between low and high grade soft tissue sarcomas.