The Multi-View Paradigm Shift in MRI Radiomics: Predicting MGMT Methylation in Glioblastoma
📝 Abstract
Non-invasive inference of molecular tumor characteristics from medical imaging is a central goal of radiogenomics, particularly in glioblastoma (GBM), where O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation carries important prognostic and therapeutic significance. Although radiomics-based machine learning methods have shown promise for this task, conventional unimodal and early-fusion approaches are often limited by high feature redundancy and an incomplete modeling of modality-specific information. In this work, we introduce a multi-view latent representation learning framework based on variational autoencoders (VAE) to integrate complementary radiomic features derived from post-contrast T1-weighted (T1Gd) and Fluid-Attenuated Inversion Recovery (FLAIR) magnetic resonance imaging (MRI). By encoding each modality through an independent probabilistic encoder and performing fusion in a compact latent space, the proposed approach preserves modality-specific structure while enabling effective multimodal integration. The resulting latent embeddings are subsequently used for MGMT promoter methylation classification.
💡 Analysis
Non-invasive inference of molecular tumor characteristics from medical imaging is a central goal of radiogenomics, particularly in glioblastoma (GBM), where O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation carries important prognostic and therapeutic significance. Although radiomics-based machine learning methods have shown promise for this task, conventional unimodal and early-fusion approaches are often limited by high feature redundancy and an incomplete modeling of modality-specific information. In this work, we introduce a multi-view latent representation learning framework based on variational autoencoders (VAE) to integrate complementary radiomic features derived from post-contrast T1-weighted (T1Gd) and Fluid-Attenuated Inversion Recovery (FLAIR) magnetic resonance imaging (MRI). By encoding each modality through an independent probabilistic encoder and performing fusion in a compact latent space, the proposed approach preserves modality-specific structure while enabling effective multimodal integration. The resulting latent embeddings are subsequently used for MGMT promoter methylation classification.
📄 Content
Over the past two decades, machine learning (ML) has evolved from unimodal and single-view paradigms toward integrative approaches that combine information from multiple sources, representations, or tasks [1]. One of the earliest manifestations of this integrative thinking was ensemble learning, which merges the outputs of several classifiers or regressors to achieve greater accuracy and robustness than any individual model [2][3][4]. Classic ensemble techniques such as bagging, boosting, and stacking demonstrated how model diversity could be systematically exploited, setting an important precedent for later developments in multi-view, multimodal, and multi-task learning [5,6].
Building on these foundations, the ML community extended the “multi-concept” to encompass not only the fusion of models but also the integration of diverse data representations (multiview), heterogeneous modalities (multimodal), and related learning objectives (multi-task). A prominent example arises in medical imaging, where modalities such as MRI inherently provide multiple complementary perspectives through sequences like T1-weighted (T1), T2-weighted (T2), T1Gd, and FLAIR [7]. This diversity exemplifies the need for integrative learning frameworks capable of leveraging correlated but distinct views of the same underlying anatomy or pathology [8,9]. Such integrative approaches are particularly relevant to clinical decision-making, which is naturally multimodal -radiological findings are interpreted alongside patient history, laboratory results, histopathology, and genomic data. Advances in multi-view and multimodal ML have therefore enabled the joint analysis of diverse MRI sequences and clinical variables, improving diagnostic accuracy and generalizability even when datasets are incomplete or heterogeneous [10,11].
A key example of this trend is radiogenomics, an emerging domain that correlates quantitative imaging phenotypes with underlying genomic alterations, enabling the non-invasive inference of molecular biomarkers [12,13]. Of these biomarkers, the MGMT promoter methylation status is among the most critical prognostic and predictive markers in GBM [14,15]. According to the World Health Organisation (WHO) 2021 Classification of Tumors of the Central Nervous System (CNS), GBM is defined as an adult-type diffuse glioma, Isocitrate Dehydrogenase (IDH)-wildtype, CNS WHO grade 4-the most aggressive astrocytic malignancy [16]. This classification underscores the clinical importance of molecular markers such as MGMT, which contribute to therapeutic decision-making and stratification of GBM patients. Since conventional determination requires invasive biopsy and molecular testing, recent research has focused on radiomic strategies for non-invasive prediction of MGMT methylation [17][18][19][20]. These methods exploit the complementary strengths of different MRI sequences (multi-view integration) and combine imaging features with clinical or molecular data (multimodal fusion) to capture tumor heterogeneity and underlying biological complexity.
In this work, we demonstrate that learning modality-aware latent representations from complementary MRI-derived radiomics improves non-invasive prediction of MGMT promoter methylation compared to classical unimodal and early-fusion approaches. To systematically assess this advantage, we compare unimodal radiomics, classical multimodal radiomics, and a multi-view variational autoencoder framework within a unified and methodologically consistent experimental setting.
In 2016, the FAIR Guiding Principles for scientific data management and stewardship were published in Scientific Data [21] defining a framework to improve the Findability, Accessibility, Interoperability, and Reuse (FAIR) of digital assets. The principles emphasize machineactionability, enabling computational systems to efficiently find, access, and reuse data with minimal human intervention, addressing the growing scale and complexity of scientific data 1 . These principles have been widely adopted across biomedical and imaging research. The present study follows the FAIR framework, utilizing open-access and interoperable imaging data to ensure transparency and reproducibility. In Europe, infrastructures such as Euro-including the enhancing core, necrotic core, and peritumoral edema, which serve as spatial references for feature computation and region-specific analysis.
Beyond imaging, the dataset incorporates clinical and molecular metadata, such as MGMT promoter methylation status, IDH1 mutation, and overall survival, enabling integrative radiogenomic studies that link imaging-derived features with molecular characteristics. This combination of harmonized multimodal imaging, expert annotations, and molecular profiling makes the UPenn-GBM collection a robust benchmark for developing and validating Artificial Intelligence (AI)-based GBM prediction models.
In this study, the UPenn-GBM dataset forms the core of the experimental f
This content is AI-processed based on ArXiv data.