Abstract
Question: Can deep learning (DL) based radiomics using routine MRI noninvasively predict isocitrate dehydrogenase (IDH) mutation and 1p/19q co-deletion status in gliomas, and what factors affect diagnostic accuracy?
Findings: This meta-analysis of 72 studies demonstrated 80% sensitivity and 85% specificity for predicting IDH mutation, and 75% sensitivity and 82% specificity for 1p/19q co-deletion status.
Clinical Relevance: MRI-based DL models show clinically useful accuracy for noninvasive molecular classification. However, clinical adoption requires standardized automated segmentation, data harmonization, and multi-center external validation.
Introduction
Gliomas are the most common primary central nervous system tumors, and their clinical management is increasingly dictated by molecular markers—specifically IDH mutation and 1p/19q co-deletion. Currently, identifying these markers requires invasive surgical biopsy, which carries risks and may not account for intratumoral heterogeneity.
While MRI is the standard non-invasive tool for monitoring, the radiographic patterns associated with these genetic mutations are often invisible to the human eye. Deep learning-based radiomics aims to extract these "sub-visual" features. This systematic review and meta-analysis evaluates whether these models are ready for clinical integration.
Methods
The authors performed a PRISMA-guided systematic review, screening over 1,500 publications and ultimately including 104 studies for qualitative synthesis and 72 for quantitative meta-analysis.
Quality and Statistical Framework
- Quality Benchmarking: Studies were graded using the Radiomics Quality Score (RQS) and a modified QUADAS-2 tool tailored for AI (incorporating CLAIM checklist elements).
- Statistical Modeling: A bivariate random effects model was employed to pool sensitivity and specificity, accounting for the inherent trade-offs across different model thresholds.
- Meta-regression: Researchers used meta-regression to identify drivers of performance variability, such as segmentation methods and the level of DL integration.
Data Landscape
Research in this field is heavily concentrated in China and the USA. While 35% of studies used private in-house collections, many relied on public benchmarks like the Cancer Imaging Archive (TCIA), Brain Tumor Segmentation (BraTS) challenge, and the Erasmus Glioma Database (EGD).
Key Results and Insights
The meta-analysis yielded strong diagnostic performance metrics for test cohorts:
- IDH Mutation Prediction: 80% Sensitivity (95% CI: 77–83%), 85% Specificity (95% CI: 81–87%), and an AUC of 0.88.
- 1p/19q Co-deletion Prediction: 75% Sensitivity (95% CI: 65–82%), 82% Specificity (95% CI: 75–88%), and an AUC of 0.85.
Technical Observations
- The End-to-End Advantage: Models using "end-to-end" DL (learning features directly from pixels) generally outperformed hybrid models that used DL merely for feature extraction.
- Segmentation as a Bottleneck: Automated and semi-automated segmentation methods provided more consistent results than manual outlines, which were a primary driver of inter-study variability.
- Architecture Evolution: The field is shifting from standard Convolutional Neural Networks (CNNs) toward Transformers and hybrid architectures with attention mechanisms to better capture global context.
Discussion and Future Directions
Despite promising results, several hurdles remain for clinical adoption. The "domain shift"—variability in MRI appearance across different scanners—requires standardized data harmonization techniques. Furthermore, the "black box" nature of DL models necessitates improved interpretability for clinician trust.
The authors advocate for the development of "Foundation Models" in neuro-oncology—large-scale models trained on diverse, multi-center data—and emphasize that prospective clinical trials are the essential next step for validation.
Reference: Eur Radiol (2025) Farahani S, Hejazi M, Tabassum M, et al. DOI: 10.1007/s00330-025-11898-2
Summary updated