ABSTRACT
Objective To explore the classification of gliomas according to the theory and method of radiomics. Methods In this study, 161 pathologically confirmed glioma patients were retrospectively selected from 2012 to 2016 including 52 low-grade gliomas and 109 high-grade gliomas.Three hundred and forty-six quantization features were extracted from the MRI images, including shape, density, texture and wavelet imaging features. Mutual information and logistic regression model were used to select feature reduction and prediction model. The predictive ability of the model was validated using 10-fold cross-validation. Results Nineteen radiomics features were chosen from 346 quantization features. The sensitivity of the model was 96.3% (105/109), the specificity was 78.8% (41/52), the area under the curve (AUC) was 0.952 7, and the accuracy was 90.7%(146/161). Conclusion The solution proposed in this paper showed that radiomics can non-invasively and quickly provide an adjunct to the clinical grade of glioma with high accuracy.