Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis

Foundation models are changing the way we develop medical artificial intelligence. By learning broadly generalizable features across diverse data modalities, a single model can be rapidly adapted to address multiple modalities and tasks with minimal supervision. This potential comes with the urgent need to reliably benchmark, understand and compare the performance and clinical impact of foundation models across data modalities and clinical tasks. We introduce UNICORN, a fundamentally new benchmarking concept for medical foundation models. UNICORN brings four main contributions to medical artificial intelligence. First, a framework that enables a one-to-many benchmarking approach, where a single foundation model is tested across multiple tasks and data modalities. Here, we populate it with 20 tasks across radiology, pathology, and clinical text, covering classification, detection, segmentation, regression, and vision-language generation. Second, a publicly available evaluation platform that implements, for the first time, a two-step approach to run foundation models for data encoding followed by custom task-specific adaptation via few-shot learning and linear probing mechanisms. Third, we create a meta-model that combines state-of-the-art foundation models in pathology, radiology and language with novel task-specific adapters that address all UNICORN tasks, which we refer to as Unicorn Model-0 (UM-0). Finally, we design a novel UNICORN score to benchmark and compare model performance across all tasks. We present the results of UM-0 using sequestered test data from over 2,400 patients, 3,700 vision cases, and 2,400 clinical reports from 17 institutions across eight countries, spanning eight anatomical regions and four imaging modalities. Data, baselines, and evaluation platform are publicly accessible at unicorn.grand-challenge.org.

Building foundation models for medical imaging requires pooling data across institutions, yet privacy regulations prohibit centralized aggregation. Existing Federated Foundation Models either fine-tune natural-image models with poor medical-domain transfer, or train from scratch within a single modality, lacking the flexibility to unify tasks. We identify an under-explored challenge, Imaging Modality Heterogeneity, where clients operate under two structural regimes: Overlapped (shared modalities with heterogeneous label distributions) and Non-overlapped (fully disjoint modalities per client). We propose FM$^2$, a unified framework that trains the core backbone from scratch to preserve medical domain fidelity while optionally incorporating biomedical pretrained encoders for vision-language alignment. FM$^2$ equips each client with dual Mixture-of-Experts modules (a Class-wise MoE for personalized category knowledge and a Domain-wise MoE for shared cross-modality representations), coupled with a Heterogeneous Modality Alignment (HMA) regularizer that explicitly aligns modality-specific expert parameters, admitting provable $O(1/\sqrt{T})$ convergence and generalization guarantees. FM$^2$ further incorporates Caption-Enhanced Learning (CEL), where locally retained GPT-4o-generated captions serve as a textual semantic bridge enabling representation transfer across clients with disjoint modalities, and demonstrates extensibility to Federated Medical VQA. Experiments on our MIMH benchmark (classification and CEL) and real-world medical VQA datasets confirm consistent superiority over state-of-the-art federated baselines and strong out-of-modality generalization across all three tasks.

Multimodal models in oncology can produce accurate predictions, but accurate prediction does not reveal whether the model has learned biology that is shared across modalities, biology confined to one modality, or spurious correlations that reflect confounders rather than genuine biology. We introduce DECAT, a model-agnostic post-hoc evaluation framework that classifies multimodal representations into four diagnostic scenarios for a given task and modality, using five null-referenced metrics and a rule-based decision procedure. The framework operates on learned representations, requires no knowledge of which specific confounder is present, and returns indeterminate when the evidence is insufficient. We validate DECAT on synthetic data across four multimodal model classes (over 2,500 trained representations) and on real data from 8,979 TCGA patients, evaluating both multimodal embeddings and five pretrained pathology foundation models. Entangled models (e.g., CLIP) achieve near-perfect shared biology detection but falsely claim shared biology in the majority of cases where it is absent on real foundation model embeddings. This false claim rate increases with confound strength so that larger cohorts and stronger representations produce more confident but still incorrect diagnoses. Applied to both multimodal TCGA embeddings and five pathology foundation models without paired RNA, DECAT detects confounding invisible to AUROC without requiring the confounder labels, as confirmed by post-hoc stratification.