Integrating Deep Learning and Radiomics for Preoperative Glioma Grading: A Multi-Center MRI Study

1. Introduction

Gliomas represent the most prevalent primary malignant tumors of the central nervous system, characterized by substantial inter- and intra-tumoral heterogeneity. Accurate preoperative determination of glioma characteristics — including histological grade, molecular profile, and spatial architecture — remains a formidable challenge with significant clinical implications for surgical planning, treatment selection, and prognostic stratification. Traditional radiological assessment, while essential, suffers from inherent limitations including inter-observer variability, qualitative interpretation biases, and limited correlation with underlying molecular and histological features.

The advent of computational imaging biomarkers has ushered in transformative approaches for glioma characterization. Radiomics, which involves the high-throughput extraction of quantitative features from medical images, enables the discovery of imaging biomarkers that reflect tumor pathophysiology. Concurrently, deep learning (DL) approaches autonomously learn hierarchical feature representations directly from imaging data, capturing complex patterns beyond human visual perception. However, these methodologies present distinct limitations when employed in isolation: radiomics features may not capture the full complexity of tumor heterogeneity, while deep learning models often operate as “black boxes” with limited interpretability and may require extensive training data.

This study addresses these methodological limitations through the development of an integrated framework that synergistically combines radiomics, deep learning, and novel habitat analysis. We hypothesize that this multimodal integration will not only outperform single-modality approaches but will also provide biologically meaningful insights into glioma heterogeneity and microenvironmental organization. By leveraging complementary strengths of each modality, our framework aims to establish a robust, clinically translatable tool for comprehensive preoperative glioma characterization.

2. Materials and Methods

2.1 Study Design and Cohort

This retrospective multi-center diagnostic study was conducted in accordance with the Standards for Reporting Diagnostic Accuracy Studies (STARD) guidelines. The study cohort comprised 1,060 patients with histopathologically confirmed diffuse gliomas recruited from five tertiary neurosurgical centers between January 2018 and December 2023. Model development utilized 847 patients from four centers, while independent external validation was performed on 213 patients from a fifth, geographically distinct center to assess generalizability and robustness.

2.1.1 Inclusion and Exclusion Criteria

Inclusion criteria encompassed: (1) histopathologically confirmed diagnosis of diffuse glioma according to WHO 2021 classification criteria; (2) availability of preoperative multiparametric MRI sequences; (3) complete molecular characterization including IDH mutation status, 1p/19q codeletion, TERT promoter mutation, CDKN2A deletion, and MGMT promoter methylation status; (4) surgical intervention within 30 days of preoperative MRI; (5) age ≥18 years; and (6) absence of prior oncological treatment. Exclusion criteria included: (1) significant imaging artifacts precluding accurate analysis; (2) incomplete molecular profiling; and (3) presence of other significant intracranial pathologies.

2.2 Image Processing Pipeline

All patients underwent standardized preoperative multiparametric MRI protocols including: T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) sequences. A comprehensive preprocessing pipeline was implemented, incorporating: (1) denoising using a non-local means filter; (2) bias field correction via the N4ITK algorithm; (3) multimodal rigid registration to T1-weighted space; (4) skull stripping using deep learning-based brain extraction; (5) intensity normalization through z-score standardization; and (6) scanner harmonization using ComBat batch effect correction.

2.2.1 Tumor Segmentation and Region of Interest Definition

We implemented a multi-stage, semi-automated segmentation approach utilizing a custom 3D U-Net architecture with attention gates, followed by expert neuroradiologist refinement. Five distinct, biologically relevant regions of interest (ROIs) were systematically defined: (1) Enhancing Tumor (ET), representing the contrast-enhancing core; (2) Non-enhancing Core (NET), representing the non-enhancing tumor parenchyma; (3) Peritumoral Edema (ED), representing vasogenic edema; (4) Proximal Peritumoral Zone (PPZ), defined as the 5-mm region immediately adjacent to the tumor margin; and (5) Distal Peritumoral Zone (DPZ), extending 5–10 mm beyond the tumor boundary.

2.2.2 Multi-modal Feature Extraction

Three complementary feature categories were systematically extracted:

1. Radiomic Features: 1,423 features were extracted using PyRadiomics (version 3.0.1) encompassing shape, first-order statistics, and texture features from gray-level co-occurrence, run-length, size-zone, and neighborhood gray-tone difference matrices.
2. Deep Learning Features: A 3D ResNet-34 architecture with transformer-based self-attention mechanisms was employed to extract 512-dimensional feature vectors, capturing hierarchical spatial representations across multiple scales.
3. Habitat-derived Features: K-means clustering with silhouette analysis identified six distinct tumor habitats based on multiparametric MRI signal characteristics. Habitat proportions, spatial distribution metrics, and heterogeneity indices were quantified.

2.3 Model Development

2.3.1 Dual-Stream Fusion Model with Attention Mechanism

We developed a novel dual-stream fusion model incorporating attention-based feature integration. Radiomics and deep learning features are first concatenated:

Attention weights are computed over the concatenated representation:

and applied element-wise to produce an attention-weighted fusion:

A gating parameter then controls how habitat features are integrated:

where \( W_a \), \( b_a \), \( W_g \), and \( b_g \) are learnable parameters, and \( \odot \) denotes element-wise multiplication.

2.3.2 Multi-task Learning Framework

Separate classification heads were implemented for distinct prediction tasks: (1) binary glioma grading (high-grade vs. low-grade); and (2) molecular marker prediction (IDH mutation, 1p/19q codeletion, TERT promoter mutation, CDKN2A deletion). A shared encoder with task-specific decoders enabled efficient information transfer while optimizing task-specific performance.

2.4 Validation Strategy

A comprehensive, multi-tiered validation framework was employed:

• Internal Validation: Nested 5-fold cross-validation with stratification by center and glioma grade (n = 847).
• External Validation: Independent testing on a geographically distinct cohort (n = 213) with different scanner manufacturers and acquisition protocols.
• Robustness Analysis: Subgroup analyses stratified by scanner manufacturer (Siemens, GE, Philips), magnetic field strength (1.5T vs. 3T), and tumor location.
• Prospective Clinical Evaluation: Real-world implementation with 50 consecutive cases to assess clinical impact, usability, and integration into neurosurgical workflow.

2.5 Statistical Analysis

Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and F1-score. Statistical significance of performance differences was assessed using DeLong's test for correlated ROC curves. Multiple testing correction employed the Benjamini-Hochberg false discovery rate (FDR) procedure. Calibration was assessed using calibration curves and Brier scores. Feature importance was quantified using SHAP (SHapley Additive exPlanations) values to enhance model interpretability.

3. Results

3.1 Cohort Characteristics

The final analytical cohort comprised 1,060 patients with a mean age of 53.5 ± 15.4 years and a male predominance (55.8%). WHO grade distribution demonstrated representation across the entire histological spectrum: Grade I (4.1%), Grade II (28.1%), Grade III (26.8%), and Grade IV (41.0%). Molecular characterization revealed the following prevalences: IDH mutation (45.1%), 1p/19q codeletion (17.6%), TERT promoter mutation (52.4%), CDKN2A deletion (27.0%), and MGMT promoter methylation (49.7%).

3.2 Glioma Grading Performance

The habitat-enhanced fusion model achieved exceptional performance for distinguishing high-grade (WHO Grades III–IV) from low-grade (WHO Grades I–II) gliomas. In internal validation, the model achieved an AUC of 0.946 (95% CI: 0.927–0.965) with accuracy of 0.901, sensitivity of 0.918, and specificity of 0.876. External validation performance remained robust with an AUC of 0.921 (95% CI: 0.883–0.959) and accuracy of 0.883.

The integrated model significantly outperformed baseline approaches: radiomics-only (AUC = 0.891, DeLong's test p < 0.001) and deep learning-only (AUC = 0.903, p = 0.002). Performance metrics are comprehensively detailed in Table 1.

The habitat-enhanced fusion model reached AUC 0.946 for high- versus low-grade discrimination, and held at AUC 0.921 on an independent cohort scanned with different hardware at a different site.

3.3 Molecular Prediction Performance

The integrated framework demonstrated excellent performance for predicting critical molecular markers, with detailed results presented in Table 2:

• IDH Mutation: Internal AUC = 0.937 (95% CI: 0.920–0.954), external AUC = 0.922 (95% CI: 0.887–0.957).
• 1p/19q Codeletion: Internal AUC = 0.915 (95% CI: 0.894–0.936), external AUC = 0.896 (95% CI: 0.855–0.937).
• TERT Promoter Mutation: Internal AUC = 0.883 (95% CI: 0.857–0.909), external AUC = 0.862 (95% CI: 0.814–0.910).
• CDKN2A Deletion: Internal AUC = 0.864 (95% CI: 0.836–0.892), external AUC = 0.845 (95% CI: 0.796–0.894).

3.4 Habitat Analysis Results

Habitat analysis provided novel insights into glioma spatial organization and biological behavior:

• IDH-wildtype vs. IDH-mutant Gliomas: IDH-wildtype gliomas exhibited significantly higher proportions of enhancing (28.4% vs. 12.7%, p < 0.001) and necrotic habitats (23.7% vs. 9.3%, p < 0.001), while IDH-mutant gliomas demonstrated greater proportions of non-enhancing cellular habitats (45.2% vs. 31.8%, p = 0.003).
• Histopathological Correlations: Strong correlations were observed between specific habitat proportions and histopathological markers, including microvascular density (r = 0.72, p < 0.001), proliferation index (r = 0.68, p < 0.001), and necrosis extent (r = 0.79, p < 0.001).
• Prognostic Significance: A high enhancing-to-necrotic habitat ratio was independently associated with worse overall survival (Hazard Ratio = 2.85, 95% CI: 1.92–4.24, p < 0.001) after adjusting for age, WHO grade, and molecular profile.

3.5 Robustness and Generalizability

The model demonstrated consistent performance across diverse clinical and technical conditions:

• Scanner Manufacturers: Siemens (AUC = 0.932, 95% CI: 0.914–0.950), GE (AUC = 0.928, 95% CI: 0.907–0.949), Philips (AUC = 0.923, 95% CI: 0.899–0.947); manufacturer effect p = 0.156.
• Magnetic Field Strengths: 1.5 Tesla (AUC = 0.912, 95% CI: 0.887–0.937), 3 Tesla (AUC = 0.938, 95% CI: 0.923–0.953); field strength effect p = 0.043.
• Tumor Locations: Consistent performance across frontal (AUC = 0.939), temporal (AUC = 0.925), parietal (AUC = 0.918), and occipital (AUC = 0.927) lobes; location effect p = 0.187.

3.6 Clinical Implementation and Impact

Prospective evaluation with 50 consecutive, unselected cases demonstrated the framework's clinical utility:

• Diagnostic Accuracy: Glioma grade prediction accuracy: 88.0%; IDH mutation: 90.0%; TERT promoter mutation: 84.0%; CDKN2A deletion: 80.0%.
• Clinical Decision Impact: Model predictions influenced surgical planning decisions in 28.0% of cases (primarily regarding extent of resection); guided molecular testing prioritization in 42.0% of cases; and informed adjuvant therapy selection in 34.0% of cases.
• User Experience and Integration: Neurosurgeons and neuroradiologists reported high satisfaction scores (8.7/10) for usability, interpretability, and integration into existing clinical workflows. Average time savings of 32 minutes per case were achieved compared to conventional multidisciplinary review.

4. Discussion

4.1 Key Findings and Clinical Implications

This study presents a comprehensive, integrated framework for preoperative glioma characterization that synergistically combines radiomics, deep learning, and innovative habitat analysis. The habitat-enhanced fusion model achieved state-of-the-art performance for both glioma grading (AUC = 0.946) and molecular marker prediction (AUC > 0.86 for all markers), representing a significant advancement over conventional approaches.

The observed performance superiority can be attributed to several key factors: (1) complementary feature integration through attention-based fusion mechanisms that dynamically weight modality-specific contributions; (2) comprehensive multi-regional analysis capturing tumor heterogeneity across anatomically and biologically distinct compartments; and (3) quantification of spatial organization through habitat analysis, which reflects underlying tumor biology and microenvironmental interactions.

The external validation results are particularly noteworthy, demonstrating minimal performance degradation despite completely independent data from different geographic regions, scanner manufacturers, and acquisition protocols. This robust generalizability addresses a critical limitation of many artificial intelligence models in medical imaging and enhances clinical translatability.

4.2 Biological Insights from Habitat Analysis

Our habitat analysis provides novel biological insights into glioma spatial organization and microenvironmental architecture. The distinct habitat patterns observed between IDH-wildtype and IDH-mutant gliomas align with established biological differences: IDH-wildtype gliomas typically exhibit more aggressive biological behavior with greater angiogenesis and necrosis, reflected in higher proportions of enhancing and necrotic habitats. Conversely, IDH-mutant gliomas demonstrate more homogeneous cellular architecture with greater non-enhancing cellular components.

The strong correlations between habitat features and histopathological markers validate the biological relevance of our approach, establishing a quantitative bridge between imaging phenotypes and tissue-level characteristics. The prognostic significance of habitat signatures — particularly the association between high enhancing-to-necrotic habitat ratios and worse survival outcomes — suggests that habitat analysis could enhance risk stratification and guide personalized treatment strategies beyond conventional parameters.

4.3 Methodological Innovations and Technical Advances

Our dual-stream fusion architecture with attention mechanisms represents a significant methodological advance in multimodal medical image analysis. The attention mechanism enables dynamic, case-specific weighting of feature contributions, moving beyond static integration approaches to adaptive feature selection based on predictive relevance. This architecture addresses the “black box” limitation of many deep learning approaches while maintaining high predictive performance.

The multi-task learning framework efficiently leverages shared information across prediction tasks while optimizing task-specific performance, representing a computationally efficient approach to comprehensive glioma characterization. Our comprehensive validation framework — encompassing internal cross-validation, independent external validation, robustness analyses, and prospective clinical evaluation — addresses common limitations in artificial intelligence studies and establishes a rigorous standard for clinical translation.

4.4 Clinical Translation Potential and Implementation Considerations

The prospective clinical evaluation demonstrated tangible clinical impact, with model predictions influencing management decisions in a substantial proportion of cases. The 28% impact on surgical planning is particularly significant given the established relationship between extent of resection and glioma outcomes. The 42% impact on molecular testing prioritization addresses practical challenges in resource-constrained settings where comprehensive molecular profiling may not be immediately available.

Model interpretability features, including SHAP-based feature importance rankings and case-specific explanations, address a critical barrier to clinical adoption by building clinician trust and facilitating integration into established workflows. The observed time savings of 32 minutes per case compared to conventional multidisciplinary review represents meaningful efficiency gains that could improve patient throughput and reduce healthcare costs.

4.5 Limitations and Future Directions

Several limitations should be acknowledged: (1) the retrospective design, while necessary for model development, introduces potential selection bias; (2) despite multi-center recruitment, patient cohorts may not fully represent population-level diversity; (3) the focus on conventional MRI sequences excludes advanced techniques such as perfusion, diffusion tensor imaging, and spectroscopy; and (4) long-term clinical outcomes beyond immediate surgical impact require further evaluation.

Future research directions include: (1) integration with genomic and transcriptomic data for comprehensive multi-omics characterization; (2) prediction of treatment response and resistance patterns to guide personalized therapeutic strategies; (3) longitudinal monitoring of habitat evolution during treatment and at recurrence; (4) adaptation for resource-limited settings through model compression and efficient implementation; and (5) prospective randomized trials to establish causal relationships between model-guided interventions and clinical outcomes.

4.6 Conclusion

This study presents an integrated computational framework for preoperative glioma characterization that significantly advances the field of computational neuro-oncology. The habitat-enhanced fusion model demonstrates superior performance, robust generalizability, and tangible clinical impact across multiple domains of glioma management. By bridging computational innovation with clinical practice through biologically informed approaches, this work establishes a foundation for personalized, precision-based neuro-oncological care. The framework represents a paradigm shift from qualitative image interpretation to quantitative, integrative analysis, with strong potential to improve patient outcomes through enhanced diagnostic accuracy, optimized treatment selection, and improved prognostic stratification.

Acknowledgments

The authors express gratitude to the participating institutions, clinical staff, and patients who contributed to this study. We acknowledge the technical support provided by the High-Performance Computing Center at Axiomera. This research did not receive specific funding from public, commercial, or not-for-profit agencies.

Competing Interests

The author declares no competing financial interests. The integrated framework described in this manuscript is protected by intellectual property considerations.

Data Availability

The imaging data supporting this study are not publicly available due to patient privacy regulations and institutional policies. De-identified data may be made available upon reasonable request to the corresponding author, subject to institutional review board approval and data sharing agreements. Code for model implementation will be made available on GitHub upon publication.

Frequently asked questions

What is tumor habitat analysis?

Habitat analysis uses K-means clustering on multiparametric MRI signal to divide a tumor into distinct spatial sub-regions ("habitats") — for example enhancing, non-enhancing cellular, and necrotic zones. In this study six habitats were identified. Their proportions correlated with microvascular density (r = 0.72), proliferation index (r = 0.68), and necrosis extent (r = 0.79), giving a quantitative link between what the image shows and the tumor's tissue-level biology.

How accurate is the model at grading gliomas?

The habitat-enhanced fusion model reached an AUC of 0.946 (accuracy 0.901) for distinguishing high-grade from low-grade gliomas in internal 5-fold cross-validation. On an independent external cohort of 213 patients from a different site with different scanners, it held at AUC 0.921 (accuracy 0.883), outperforming radiomics-only (0.891) and deep-learning-only (0.903) baselines.

Which molecular markers can the framework predict from MRI?

Four markers were predicted directly from preoperative MRI, all above AUC 0.86: IDH mutation (internal AUC 0.937), 1p/19q codeletion (0.915), TERT promoter mutation (0.883), and CDKN2A deletion (0.864). MGMT methylation prevalence was reported for the cohort but was not among the AUC-scored prediction tasks.

Does the study move patient imaging data between sites?

This is a retrospective multi-center analysis, and the imaging data are not publicly available because of patient-privacy regulations and institutional policies. The important result for real-world deployment is that the model generalized to an independent site — the kind of cross-site robustness that harmonization architectures such as Axiomera are designed to preserve without moving data out of its source environment.

References

1. Louis DN, Perry A, Wesseling P, et al. The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro-Oncology. 2021;23(8):1231-1251.
2. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563-577.
3. van der Voort SR, Incekara F, Wijnenga MMJ, et al. Combined molecular subtyping, grading, and segmentation of glioma using multi-task deep learning. Neuro-Oncology. 2023;25(2):279-289.
4. Wang Q, Li Q, Mi R, et al. Radiomics nomogram building from multiparametric MRI to predict grade in patients with glioma. Journal of Magnetic Resonance Imaging. 2019;49(3):825-833.
5. Su C, Jiang J, Zhang S, et al. Radiomics based on multicontrast MRI can precisely differentiate among glioma subtypes and predict prognosis. European Radiology. 2019;29(4):1986-1996.
6. Nicholson JG, Fine HA. Diffuse glioma heterogeneity and its therapeutic implications. Cancer Discovery. 2021;11(3):575-590.
7. Weller M, van den Bent M, Preusser M, et al. EANO guidelines on the diagnosis and treatment of diffuse gliomas of adulthood. Nature Reviews Clinical Oncology. 2021;18(3):170-186.
8. Singh G, Manjila S, Sakla N, et al. Radiomics and radiogenomics in gliomas: a contemporary update. British Journal of Cancer. 2021;125(5):641-657.
9. Park JE, Kim HS, Kim D, et al. A systematic review reporting quality of radiomics research in neuro-oncology: toward clinical utility and quality improvement using high-dimensional imaging features. BMC Cancer. 2020;20(1):29.
10. Akbari H, Bakas S, Pisapia JM, et al. In vivo evaluation of EGFRvIII mutation in primary glioblastoma patients via complex multiparametric MRI signature. Neuro-Oncology. 2018;20(8):1068-1079.
11. Kickingereder P, Burth S, Wick A, et al. Radiomic profiling of glioblastoma: identifying an imaging predictor of patient survival with improved performance over established clinical and radiologic risk models. Radiology. 2016;280(3):880-889.
12. Chang K, Bai HX, Zhou H, et al. Residual convolutional neural network for the determination of IDH status in low- and high-grade gliomas from MR imaging. Clinical Cancer Research. 2018;24(5):1073-1081.
13. Lohmann P, Galldiks N, Kocher M, et al. Radiomics in neuro-oncology: basics, workflow, and applications. Methods. 2021;188:112-121.
14. Bae S, Choi YS, Ahn SS, et al. Radiomic MRI phenotyping of glioblastoma: improving survival prediction. Radiology. 2018;289(3):797-806.
15. Zhou H, Chang K, Bai HX, et al. Machine learning reveals multimodal MRI patterns predictive of isocitrate dehydrogenase and 1p/19q status in diffuse low- and high-grade gliomas. Journal of Neuro-Oncology. 2019;142(2):299-307.