Robust Feature Selection Method of Radiomics for Grading Glioma

The accuracy of glioma segmentation is significantly affected by the radiomics-based prediction model for grading glioma. This study proposed a robust feature selection method that can select stable and insensitive features to the segmentation of the region of interest (ROI). The method consists of three main steps. First, stable features are selected from 360 radiomics features based on the Pearson correlation coefficient. Then, an unsupervised K-means algorithm is introduced to remove redundant features from those selected in the first step and obtain sets of K group candidate features. Finally, by using these K group feature sets to train four prediction models, the final feature set and final prediction models that have the best prediction performance are selected. Experiments were conducted on 156 glioma samples from Henan Provincial People’s Hospital between 2012 and 2016, and 11 robust features were selected. The results demonstrated excellent accuracy, sensitivity, specificity, and AUC (0.88, 0.94, 0.88, and 0.85, respectively). Compare with the performance without feature selection, a 5% increase in accuracy, sensitivity, and AUC and 13% increase in specificity were observed. The proposed feature selection method can reduce the training time by 94.04%.