CoRe-BT: A Multimodal Breakthrough for Brain Tumor Diagnosis

In the high-stakes world of neuro-oncology, accurate brain tumor classification can mean the difference between effective treatment and therapeutic failure. Yet clinicians routinely face a frustrating reality: diagnostic evidence arrives piecemeal, with magnetic resonance imaging (MRI), histopathology slides, and pathology reports often becoming available at different times during the diagnostic process. This fragmented data landscape has long hampered artificial intelligence systems that typically require complete datasets to function optimally.

Now, a research team has introduced a solution that could transform how AI approaches brain tumor diagnosis. Published on arXiv on March 4, 2026, CoRe-BT (Cross-modal Radiology-Pathology-Text Benchmark for Brain Tumor Typing) represents the first comprehensive multimodal benchmark specifically designed to study robust learning under realistic conditions of missing clinical data.

The Clinical Challenge of Incomplete Evidence

Brain tumor typing presents unique diagnostic complexities. Gliomas—the most common primary brain tumors—exhibit tremendous heterogeneity, with subtypes requiring different treatment approaches and carrying vastly different prognoses. Current diagnostic protocols typically involve:

• Radiological assessment through multi-sequence MRI (T1, T1c, T2, FLAIR)
• Histopathological examination of tissue samples via H&E-stained whole-slide images
• Pathology reports containing textual descriptions and diagnostic conclusions

In practice, these modalities rarely arrive simultaneously. A patient might undergo MRI immediately, but pathology results could take days or weeks. This temporal mismatch creates a significant challenge for AI systems trained on complete datasets that don't reflect real-world clinical workflows.

"The gap between ideal laboratory conditions and actual clinical practice has been a major barrier to AI adoption in neuro-oncology," explains the research team behind CoRe-BT. "Our benchmark directly addresses this by providing a testbed for models that must perform with whatever data is available at decision time."

Inside the CoRe-BT Dataset

The CoRe-BT dataset comprises 310 patient cases with comprehensive clinical annotations:

• Multi-sequence brain MRI for all cases, including expert-annotated tumor masks enabling region-aware modeling
• Paired pathology data for 95 cases, featuring H&E-stained whole-slide images and corresponding pathology reports
• Standardized annotations for tumor type and grade across six clinically relevant classes
• Auxiliary learning tasks supported by tumor segmentation masks

What makes CoRe-BT particularly valuable is its intentional design to simulate real-world conditions. Researchers can test models under various scenarios: MRI-only when pathology isn't yet available, pathology-only when imaging is incomplete, or full multimodal integration when all evidence is present.

Technical Innovations and Baseline Results

The research team conducted extensive baseline experiments comparing MRI-only models against multimodal approaches. Their findings reveal several important insights:

1. Complementary modality contributions: Different tumor subtypes show varying dependence on imaging versus pathology evidence. Some gliomas are more readily identifiable through radiological features, while others require histopathological confirmation.

2. Robust fusion techniques: The benchmark enables testing of advanced fusion methods that can handle missing modalities without catastrophic performance degradation—a critical requirement for clinical deployment.

3. Representation learning opportunities: CoRe-BT supports both supervised tumor typing and self-supervised representation learning across modalities, potentially leading to more generalizable AI models.

"Our baseline experiments demonstrate not just the feasibility of multimodal fusion, but the tangible benefits of integrating complementary evidence sources," the researchers note. "Even when pathology data is available for only a subset of cases, multimodal models outperform unimodal approaches on the full patient cohort."

Clinical Implications and Future Directions

CoRe-BT arrives at a pivotal moment in medical AI development. As healthcare systems increasingly look to AI for diagnostic support, benchmarks that reflect real clinical constraints become essential for translating laboratory advances into clinical utility.

The benchmark's design addresses several practical concerns:

• Temporal realism: Models can be evaluated on their ability to provide preliminary diagnoses with available data while incorporating additional evidence as it arrives
• Clinical heterogeneity: The six tumor classes capture both common and rare glioma subtypes, ensuring relevance across patient populations
• Interpretability requirements: The multimodal nature supports explainable AI approaches that can point to specific imaging or pathology features supporting diagnostic conclusions

Looking forward, CoRe-BT could catalyze several research directions:

• Development of uncertainty-aware models that quantify diagnostic confidence based on available evidence
• Cross-modal transfer learning techniques that leverage abundant imaging data to improve performance on cases with limited pathology
• Federated learning approaches that preserve patient privacy while training on distributed clinical data

The Broader AI Benchmarking Landscape

CoRe-BT joins a growing family of specialized benchmarks emerging from the arXiv ecosystem, including GAP for general AI planning, LLM-WikiRace for reasoning evaluation, and OpenSage for cross-embodiment learning. This trend toward domain-specific, clinically grounded benchmarks represents a maturation of medical AI research—moving beyond technical novelty toward practical utility.

Unlike general AI benchmarks that measure broad capabilities, specialized medical benchmarks like CoRe-BT must balance technical rigor with clinical relevance. They need to capture the messy realities of healthcare data while providing clear evaluation metrics that translate to patient outcomes.

Conclusion: Toward Clinically Useful AI

The introduction of CoRe-BT marks a significant step toward AI systems that work with clinicians rather than requiring clinicians to work around AI limitations. By explicitly addressing the reality of incomplete diagnostic evidence, this benchmark could accelerate the development of more flexible, robust, and ultimately useful diagnostic assistants.

As the research team concludes: "CoRe-BT provides more than just another dataset—it offers a grounded testbed for advancing multimodal glioma typing in scenarios that mirror actual clinical practice. Our hope is that this will bridge the gap between AI research and clinical application, ultimately improving diagnostic accuracy and patient care."

For neuro-oncologists facing the daily challenge of interpreting fragmented evidence, and for AI researchers seeking to build clinically relevant systems, CoRe-BT represents a promising convergence of technical innovation and medical necessity. The benchmark is now publicly available, inviting the research community to build upon this foundation toward more robust, clinically integrated AI diagnostic tools.