Comprehensive Genomic Subtyping of Glioma Using Semi-Supervised Multi-Task Deep Learning on Multimodal MRI

Abstract

High grade glioma (HGG) are the most common, highly infiltrative brain tumors usually with a grim outcome with low survival. Recent comprehensive genomic profiling has greatly elucidated the molecular markers in gliomas that include the mutations in isocitrate dehydrogenase (IDH), 1p/19q co-deletion and alterations in O6 methyl-guanine methyltransferase (MGMT). Prognosis of these markers can support early clinical decision making that can consequently improve the survival outcomes. Multi-modal MRI based phenotypic tools such as radiomics based multi-variate models and state-of-art convolutional neural networks (CNNs) have shown promise in identifying these genotypes. However, current techniques do not facilitate comprehensive genomic profiling of the HGG as these are focused only on a single mutation. Moreover, the models are trained on small datasets and cannot employ unlabeled data, which is abundant as pathological labelling is invasive, expensive and inaccessible in many places. In this work, we build a semi-supervised hierarchical multi-task model that can incorporate unlabeled glioma data and learn to predict multiple molecular markers simultaneously. Our framework employs the latent space from an encoder to incorporate the unlabeled data while the hierarchical multi-task model accounts for the similarity between tasks and utilizes the shared information, resulting in inductive learning that facilitates precise delineation of IDH, MGMT, 1p/19q and grade. We applied our framework to 120 labeled, 149 semi-labeled and 48 unlabeled data using T1-contrast enhanced, T2 and FLAIR images and illustrate that our model performs with an average test accuracy of 82.35% and verified the results using task wise and modality wise ablation analysis. Moreover, the class activation maps computed from each local task branch provide clinical interpretability.