Deep Learning Guided Radiogenomic Signatures for Prognostic Stratification in Glioblastoma Multiforme

Abstract

Glioblastoma multiforme (GBM) is the most aggressive and lethal primary brain tumor, with an average survival of no more than 15 months despite advances in surgery, chemotherapy, and radiotherapy. Linking imaging phenotypes with genomic frameworks can improve personalized prognosis and treatment planning. This study develops a deep learning-based radiogenomic framework that integrates high-dimensional imaging features extracted from multiparametric MRI using a convolutional neural network (CNN) with key molecular biomarkers, including EGFR amplification, IDH mutation, and MGMT promoter methylation. A multiomics fusion module combined imaging-derived features with genomic alterations to enable stratified survival prediction. The publicly available datasets were used to train and validate the framework, i.e., TCIA and TCGA-GBM. The CNN-based radiogenomic model was more successful than the traditional radiomic and dictionary learning -based approaches, with high prognostic accuracy. Survival stratification into high- and low-risk groups showed significant differences, as confirmed by Kaplan–Meier analysis, C-index, and AUC metrics. The radiogenomic markers based on the model obtained biologically meaningful information on tumor heterogeneity and a better predictive outcome than the traditional methods. Radiogenomic signatures based on deep learning make it possible to prognosticate GBM accurately, non-invasively, and biologically in a manner that is precise, relevant, and now more useful in the field of neuro-oncology. The next step in research involves future multi-institutional validation, explainable AI integration, and adding more omics data to make prognostics more accurate and clinically applicable.