NeuroRAD-FM: A Distributionally Robust Foundation Model for Precision Neuro-Oncology

Technical Summary: NeuroRAD-FM

Problem Statement
Foundation models (FMs) for neuro-oncology hold promise for non-invasive molecular characterization and prognostication but face significant barriers to clinical utility. Current models are often limited by cross-institutional distribution shifts (site variance) caused by differences in MRI scanners, acquisition protocols, and patient populations. This leads to models that overfit site-specific signatures rather than learning biologically relevant tumor features. Furthermore, existing approaches struggle with under-represented molecular alterations due to data imbalance and often rely on single self-supervised backbones or require manual tumor segmentation during pretraining, which limits scalability and generalizability across diverse clinical environments.

Methodology
The authors introduce NeuroRAD-FM, a distributionally robust foundation model designed specifically for neuro-oncology. The methodology integrates three core components:

1. Data and Pretraining: The model is pretrained on 7,414 multi-sequence brain MRI examinations aggregated from multiple institutions (including BraTS, MU-Glioma-Post, and LUMEIRE datasets). The training utilizes four MRI sequences (T1N, T1C, T2W, T2F) and employs a 3D ResNet encoder architecture. Crucially, pretraining is segmentation-free, learning representations directly from raw imaging data without requiring manual tumor masks.
2. Group-Distributionally Robust Optimization (Group-DRO): To address site variance, the authors employ a Group-DRO framework. Instead of minimizing the average loss across all samples, the algorithm optimizes for the worst-case performance across predefined groups (e.g., imaging sites). This forces the model to learn site-invariant representations that are robust to distributional shifts, ensuring performance does not degrade on unseen hospitals or outlier imaging conditions.
3. Multi-Backbone Weighted Ensemble: Unlike models relying on a single self-supervised objective, NeuroRAD-FM pretrains a family of four complementary encoders using distinct self-supervised learning (SSL) algorithms: MoCo, DINO, MAE, and BYOL. These embeddings are combined via a weighted ensemble. The weights are determined once during validation based on proportional mean AUC and then fixed, ensuring the ensemble leverages the complementary inductive biases of contrastive, reconstruction, and self-distillation objectives while maintaining deterministic inference.

Key Contributions

• Site-Invariant Representation Learning: The integration of Group-DRO with a multi-backbone ensemble successfully learns features that are invariant to acquisition sites while preserving disease-specific biological distinctions.
• Segmentation-Free Pretraining: The model learns robust tumor-centric representations without the bottleneck of manual segmentation, enabling large-scale training on heterogeneous datasets.
• Robustness to Imbalance: The framework specifically improves performance on under-represented molecular markers (e.g., CDKN2A/2B loss, TERT mutation) where data is scarce and imbalanced.
• Unified Framework: It provides a single unified latent representation adaptable to multiple downstream tasks (segmentation, molecular prediction, survival analysis) rather than requiring multiple task-specific models.

Results
The model was evaluated on independent glioma cohorts from UCSF (n=111), UPenn (n=95), and an in-house cohort (n=292).

• Tumor Segmentation: NeuroRAD-FM significantly outperformed the baseline BrainIAC model across all tumor compartments (Enhancing Tumor, Peritumoral Edema, Necrotic Core) on glioma, meningioma, and metastasis datasets. For example, on the Enhancing Tumor (ET) for glioma, NeuroRAD-FM achieved a Dice Similarity Coefficient (DSC) of 83.59 compared to BrainIAC's 60.8.
• Molecular Biomarker Prediction: The model demonstrated superior accuracy in predicting molecular status:
  • IDH1 Mutation: AUC of 85.58 vs. 62.42 (BrainIAC).
  • MGMT Methylation: AUC of 75.59 vs. 65.36 (BrainIAC).
  • Under-represented Markers: Significant improvements were observed for CDKN2A/2B loss (AUC 80.33 vs. 65.00), EGFR amplification, TERT promoter mutation, TP53 mutation, and Ki-67 index.
• Survival Prediction: NeuroRAD-FM achieved higher concordance indices (c-index) across all three institutions compared to the baseline, with improvements ranging from 3.85 to 8.02 points depending on the cohort.
• Interpretability: Grad-CAM visualizations confirmed that the model's attention mechanisms focused on anatomically relevant tumor regions rather than background noise, supporting the biological plausibility of the learned features.

Significance and Claims
The paper claims that NeuroRAD-FM addresses the critical failure modes of current organ-agnostic and single-backbone neuro foundation models: site-bias and brittle dependence on a single inductive bias. By combining Group-DRO with a multi-backbone ensemble, the framework enables robust, site-invariant representation learning that substantially improves the prediction of both common and under-represented molecular subtypes and survival outcomes across diverse neuro-oncology cohorts.

The authors position NeuroRAD-FM as a scalable framework for clinically deployable neuro-oncologic AI, arguing that it balances the need for generalization across scanners and protocols with the necessity of preserving organ-specific biological patterns. They emphasize that while the model shows strong retrospective performance, future work must prioritize prospective, pre-registered multi-center validation to ensure clinical readiness and reproducibility.