Predicting glioma molecular subtype from diffusion anisotropy indices' distributions

Abstract

Incorporation of glioma genetic mutations, including isocitrate dehydrogenase (IDH) and telomerase reverse transcriptase (TERT), provides information on overall survival and disease course. However, such mutations are determined from a biopsy sample which represent only the biopsied region. Non–invasive tumor genotype pre diction have been studied, but they mostly focusing only on the tumor. Yet, gliomas are known to infiltrate along normal–appearing white matter (NAWM), where relevant genotype information might be available. Diffusion anisotropy indices (DAIs) and dif fusion tensor eigenvalues (DTEs), derived from diffusion tensor imaging (DTI), can be used to quantify the diffusion in the NAWM of glioma patients with varying mutation status. We hypothesize that using full–distributions of DAIs and DTEs can better represent the complex tumor effects in the NAWM in comparison to usual summary statistics. In this study, we have compared the predictive values of summary statistics, full distributions and multi–Gaussian fitting (MGF) parameters of DAIs and DTEs in the NAWM for predicting IDH–TERT subgroups, IDH and TERT mutations in 70 glioma patients. Hemispheric variations were also investigated with hemisphere differ ence distributions.The results show that, full distributions can predict tumor genotype better than standard distribution parameters, and perform better than or as well as MGF parameters. Additionally, feature selection applied to full distributions further in creased classification accuracy. IDH–TERT subgroups were best predicted with 78.6% accuracy, IDH mutation with 94.3% accuracy, and TERT mutation with 88.6% accuracy. In conclusion, full distributions are better predictors of genotype prediction. Future work will focus on increasing accuracy on a larger cohort and personalized MGF.|Keywords : Glioblastoma, Magnetic Resonance Imaging, Diffusion, Diffusion Tensor Imaging, Anisotropy Indices, Distributions, Machine Learning.