Abstract
Gliomas and Glioblastomas represent a significant portion of central nervous system (CNS) tumors associated with high mortality rates and variable prognosis. In 2021, the World Health Organization (WHO) updated its Glioma classification criteria, most notably incorporating molecular markers including CDKN2A/B homozygous deletion, TERT promoter mutation, EGFR amplification, and +7/-10 chromosome copy number changes into the grading and classification of adult and pediatric Gliomas. This review explores the relative benefits and drawbacks of computational frameworks—radiomics, deep learning, and combined approaches—in leveraging MRI data to predict these essential biomarkers non-invasively.
Introduction: The Molecular Shift in Glioma Classification
Traditionally, brain tumors were classified primarily through histological examination of tissue samples. However, the 2021 WHO Classification of Tumors of the Central Nervous System (CNS-5) established molecular markers as the "gold standard" for diagnosis and grading. This shift allows for more precise tailoring of clinical interventions but presents a significant challenge: obtaining these markers typically requires invasive brain biopsies or surgical resections.
Invasive procedures carry inherent risks and may fail to capture the full spatial heterogeneity of a tumor. Furthermore, tumors located in critical "no-go" zones, such as the brainstem, may be ineligible for biopsy. Radiogenomics has emerged as a powerful solution, using AI to extract hidden patterns from non-invasive medical imaging (MRI) that correlate with these critical molecular signatures.
Computational Frameworks in Radiogenomics
Current research into non-invasive molecular profiling generally follows three primary computational frameworks:
- Radiomics: Often referred to as the "Glass Box" approach, radiomics uses mathematical filters to extract "hand-crafted" features such as texture, shape, and intensity distribution. These features are highly interpretable, allowing clinicians to understand the morphological basis (e.g., tumor roughness or edge sharpness) of a model's prediction.
- Deep Learning: This "Black Box" approach employs Convolutional Neural Networks (CNNs) and Vision Transformers to learn features directly from raw pixels. While more powerful and capable of identifying patterns invisible to the human eye, these models often lack the direct clinical interpretability of traditional radiomics.
- Combined Approaches: Representing the "New Frontier," hybrid models utilize deep learning for automated tumor segmentation and feature extraction while retaining radiomic descriptors to maintain clinical relevance.
The general workflow involves image pre-processing (such as skull-stripping and intensity normalization), followed by feature extraction and the application of machine learning classifiers (e.g., Random Forests or SVMs) to predict genetic status.
Key Insights and Performance Metrics
Radiogenomic tools have demonstrated remarkable accuracy across various biomarkers integrated into the CNS-5 guidelines:
- IDH Mutation: This is a cornerstone of the new classification. Studies consistently report high performance, with Area Under the Curve (AUC) values ranging from 0.90 to 0.96.
- 1p/19q Codeletion: Essential for diagnosing oligodendrogliomas, models have achieved diagnostic accuracies between 85% and 88%.
- TERT Promoter Mutation: While a more challenging target, recent deep learning models have reached AUCs as high as 0.89 to 0.99 in specific cohorts.
- MGMT Methylation: A critical predictor of chemotherapy response, fusion models combining multiple MRI sequences have achieved accuracies up to 96%.
Beyond simple classification, researchers are exploring "habitat imaging"—analyzing sub-regions of the tumor (such as the necrotic core versus edema) to better understand the tumor microenvironment. This is particularly vital for identifying "molecular glioblastomas"—tumors that appear low-grade histologically but possess aggressive markers like EGFR amplification, necessitating early and aggressive intervention.
Challenges and Future Directions
Despite technical successes, several hurdles remain before these tools can be integrated into routine radiological workflows. A primary concern is the lack of standardization; features extracted by different software packages can vary significantly. Initiatives like the Image Biomarker Standardisation Initiative (IBSI) are working to address this.
Additionally, the "small data" problem persists. While deep learning thrives on large datasets, medical imaging cohorts are often limited in size, leading to risks of overfitting. Federated Learning—where models are trained across multiple institutions without sensitive patient data leaving local servers—is a promising solution to this privacy and data scarcity issue.
The ultimate goal is a seamless clinical workflow where AI provides a molecular probability score immediately following an MRI scan, enabling faster, personalized, and non-invasive cancer treatment.
Summary updated