Predicting the Gene Status and Survival Outcome of Lower Grade Glioma Patients With Multimodal MRI Features

We propose a novel class of multimodal image features based on the joint intensity matrix (JIM) to model fine-grained texture signatures in the radiomic analysis of lower grade glioma (LGG) tumors. Experiments use expanded JIM features to predict the genetic status and the survival outcome of LGG patients with preoperative T1-weighted, T1-weighted post-contrast, fluid attenuation inversion recovery (FLAIR), and T2-weighted MR images from The Cancer Imaging Archive ($\text{n}=107$ ). Texture features were extracted from regions of interest labeled by a radiation oncologist and summarized by 19 parameters. These parameters are then used to contrast mutant and wild gene type groups (i.e., IDH1, ATRX, TP53, and 1p/19q codeletion) via the Wilcoxon test, and to compare short and long survival patient groups with the Kaplan-Meier estimator. Random forest (RF) classification is employed to predict gene status (i.e., mutation or wild) and survival outcome (i.e., short or long survival), as well as to identify highly group-informative features. A subset of JIM features show statistically significant relationships with LGG gene status (i.e., in IDH1 and ATRX, with corrected p <; 0.05) and survival outcome ($\text{p}=0.0001$ , $\text{HR}=0.09$ , $\text{CI}=0.03-0.3$ ). A maximum classification AUC of 78.59% was obtained for predicting IDH1 status from combined JIM and GLCM features. Classification combining all features (i.e., volume, JIM, and GLCM) results in an AUC value of 86.79% (corrected $\text{p}=0.04$ ) in predicting short and long LGG patient survival outcomes, where JIM features are generally the most informative predictors.