Room: Track 2
Purpose: with IDH1 mutation have improved prognosis compared to wild-type IDH1 status. This study investigated T2/FLAIR MR images radiomic texture descriptors as potential biomarkers for prediction of IDH1 gene mutation status of low grade gliomas (LGG) tumors.
Methods: one TCGA-LGG MR data with known IDH1 mutation status and the available voxel of interests on the TCGA portal were selected for this study. Ninety three radiomic texture descriptors were extracted from histogram and texture maps including GLCM, GLRLM, GLSZM, NGTDM, and GLDM from T2/FLAIR modality on edema region. Two feature selections approaches including random forest (RF) and generalized linear model via penalized maximum likelihood with the LASSO (Glmnet) were applied to select the most contributory radiomic texture descriptors for predicting IDH1 gene mutation status. Two supervised (5-fold RBF-SVM ) and unsupervised (RF) classification methods were used to discriminate the LGG tumors based on their IDH1 gene mutation status. The classifiers’ hyper-parameters were estimated using grid search method.
Results: selected a smaller subset of features than RF approach (five features vs. nine features) with a minor loss of accuracy of about 2%. The selected features by glmnet approach were correlation (from GLCM), inverse difference moment normalized (from GLCM), zone percentage (from GLSZM), gray level non-uniformity (from GLSZM), run length non-uniformity (from GLRLM). Classification accuracy using supervised and unsupervised were 83.6% and 82%, respectively; the supervised and unsupervised classifiers have an area under the receiver-operating characteristic curve of 0.83 and 0.81 with precision of 95.30% and 94.48%, respectively.
Conclusion: of edema region on T2/FLAIR modality could be able to predict the IDH1 mutation status of LGG tumors. Moreover, glmnet method can select the most contributory texture features for predicting IDH1 gene mutation status compare to RF method with a minor penalty which can increase speed of computation for big datasets.