Fig. 1: Study overview.
From: Multimodal fusion of pathology and radiology foundation models for WHO 2021 glioma subtyping
a Both modalities are embedded by frozen foundation models to generate a token for each image patch. A multimodal model is trained to predict glioma subtype. b Patient-matched data is sourced from TCGA. Unmatched data from UCSF-PDGM, EGD, and EBRAINS. c Interpretability Analysis was performed via attention visualization and SHAP values. d The fusion architectures employed.
