Fig. 2: Accuracy and Matthew’s Correlation Coefficient (MCC) achieved during cross validation (CV) and on TCGA.
From: Multimodal fusion of pathology and radiology foundation models for WHO 2021 glioma subtyping

Left a, b: Modality expert is a MambaMIL network. Right c, d: Modality expert is a linear layer. Multimodal models outperform unimodal models across all settings. Mixture of Experts (MoE) shows the strongest performance relative to other multimodal approaches, especially with patch-sequence expert. Patch-sequence expert is superior to patch-mean expert. Error bars represent standard deviation.