NeuroRAD-FM

NeuroRAD-FM is a foundation model purpose-built for neuro-oncology, learning general-purpose representations of brain tumor MRI that transfer to a broad panel of clinically meaningful downstream tasks. Neuro-oncology is an unusually difficult setting for machine learning: tumors are heterogeneous, molecular subtypes are imbalanced (many alterations are rare), and imaging protocols differ substantially across hospitals, which causes models trained at one site to degrade at another. NeuroRAD-FM addresses these issues directly by combining large-scale self-supervised pretraining with distributionally robust optimization (DRO), so that the learned features remain predictive for both common and uncommon endpoints and stay reliable when deployed at new institutions.

The model was introduced in a September 2025 arXiv preprint by Moinak Bhattacharya, Angelica P. Kurtz, Fabio M. Iwamoto, Prateek Prasanna, and Gagandeep Singh, a collaboration between Stony Brook University (Department of Biomedical Informatics) and Columbia University Irving Medical Center (Departments of Radiology and Neuro-Oncology). Rather than training a bespoke classifier per task, it provides a shared imaging backbone that can be probed for molecular marker status, continuous proliferation indices, and overall survival.

By framing cross-site generalization as a robustness problem rather than a domain-adaptation afterthought, NeuroRAD-FM fits into the growing class of medical-imaging foundation models that prioritize equitable performance on underrepresented patient subgroups and rare disease variants.

#Key Features

• Distributionally robust pretraining: Applies DRO to mitigate site and class imbalance, explicitly optimizing worst-case performance so rare molecular markers and minority cohorts are not sacrificed for average accuracy.
• Multi-institutional generalization: Validated across three independent centers (UCSF, University of Pennsylvania, and Columbia/CUIMC) to demonstrate that learned representations transfer beyond the training distribution.
• Broad molecular readout: Supports common markers (MGMT, IDH1, 1p/19q, EGFR), rarer alterations (ATRX, TP53, CDKN2A/2B, TERT), continuous indices (Ki-67), and overall survival from a single shared backbone.
• Self-supervised backbone: Compares multiple self-supervised objectives (BYOL, DINO, MAE, MoCo) on a 3D ResNet-50 encoder, removing the need for scarce expert annotations during pretraining.

#Technical Details

NeuroRAD-FM uses a 3D ResNet-50 encoder pretrained with self-supervised learning on 7,414 brain tumor MRI scans aggregated from public neuro-oncology cohorts, including BraTS-GLI (pre- and post-treatment), BraTS-MEN, BraTS-MET, BraTS-PED, LUMIERE, and MU-Glioma-Post. The authors evaluate several self-supervised objectives (BYOL, DINO, MAE, MoCo) and add a distributionally robust loss to counter site and class imbalance. Downstream evaluation spans three external cohorts (UCSF, n=111; UPenn, n=95; CUIMC, n=292). DRO training raised mean balanced accuracy from 0.744 to 0.785 at CUIMC and improved survival concordance across sites (UCSF c-index 0.600→0.627; UPenn 0.647→0.672; CUIMC 0.592→0.597), with the clearest gains on underrepresented endpoints such as CDKN2A/2B, ATRX, and Ki-67.

#Applications

NeuroRAD-FM targets non-invasive characterization of brain tumors from routine MRI, where molecular status is otherwise obtained from biopsy or resection. By predicting markers like MGMT methylation, IDH1 mutation, and 1p/19q codeletion, along with proliferation indices and survival risk, it can support neuro-radiologists and neuro-oncologists in tumor stratification, treatment planning, and prognostication, especially in glioblastoma and glioma management. Its emphasis on cross-site robustness makes it particularly relevant for deployment across hospital networks with differing scanners and acquisition protocols.

#Impact

NeuroRAD-FM contributes to a wider effort to build imaging foundation models that remain trustworthy on rare classes and unseen sites rather than only maximizing average accuracy on a single cohort. Its central methodological contribution, folding distributionally robust optimization into self-supervised pretraining, offers a template for other medical-imaging domains where class and institutional imbalance are endemic. As a recent preprint, its results await peer review and independent external validation, and no public code or model weights had been released at the time of writing; reported gains, while consistent across sites, are modest in absolute terms and should be interpreted as evidence of robustness rather than a clinical-grade benchmark.