Grading glioma by radiomics with feature selection based on mutual information

Grading of glioma is crucial for both treatment decisions and prognosis assessments. This study proposes a fast, simple, and accurate prediction framework for the non-invasive grading of glioma based on radiomics. The framework consists of four main steps. First, glioma images were subjected to semi-automatic segmentation to reduce the heavy workload. Then, 346 radiomics features were calculated from the segmented regions of interest. However, selecting features directly from such a large set to train the prediction model might lead to overfitting. Therefore, a de-redundancy algorithm was proposed to construct a candidate feature set based on mutual information. Finally, feature selection was executed using elastic net, and a grading model with linear regression was built. The proposed non-invasive solution for the grading of glioma can potentially hasten treatment decision, with the use of a de-redundancy algorithm that significantly improved the prediction accuracy. Experiments were conducted on 161 glioma samples from Henan Provincial People's Hospital between 2012 and 2016, and results demonstrated the accurate grading effect and the generality of the de-redundancy algorithm. Moreover, the proposed framework exhibited desirable sensitivity (93.57%), specificity (86.53%), AUC (0.9638) and accuracy (91.30%).