Foundation Models Enhance Multimodal Cancer Analysis Trustworthiness

Foundation models (FMs) are increasingly recognized for their ability to extract powerful representations from medical data. However, their generalizability, especially to new datasets with distribution shifts, remains an area requiring further investigation. This research systematically assessed FM-based representations for computational pathology tasks using real-world commercial oncology datasets, specifically focusing on breast cancer (IH-BC) and non-small cell lung cancer (IH-NSCLC). The study analyzed two modalities: whole-slide images and transcriptomic profiles. Initial unimodal evaluations across five FMs on eight classification tasks revealed that both image and omics representations carry complementary predictive signals. This suggests that neither modality alone provides a complete picture, highlighting the potential for multimodal approaches. Further investigation into multimodal fusion strategies demonstrated that combining paired image-omics representations yields additional performance gains, particularly in scenarios where no single modality provides a dominant signal. To enhance clinical applicability, the trustworthiness of selected unimodal and multimodal pipelines was assessed using conformal prediction. This revealed that even when a model's point prediction was incorrect, the true diagnosis was often still contained within the prediction set, underscoring the value of uncertainty-aware inference for robust clinical decision support.