Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis
Researchers have systematically evaluated foundation models (FMs) for multimodal cancer analysis, focusing on their ability to generalize to out-of-distribution data. The study utilized whole-slide images and transcriptomic profiles from two commercial cohorts, IH-BC and IH-NSCLC. Findings indicate that FM representations perform competitively on unseen data, and multimodal fusion offers benefits primarily when individual modalities do not provide dominant signals. The research also highlighted the value of uncertainty-aware inference for clinical support, as conformal prediction demonstrated that true diagnoses often remain within prediction sets when point predictions fail. AI
IMPACT This research provides insights into the generalization capabilities of foundation models in medical imaging and omics data, informing their potential clinical application.
