ME-Mamba: Multi-Expert Mamba with efficient knowledge capture and fusion for multimodal survival analysis

Abstract

Multimodal survival analysis integrating Whole Slide Images (WSIs) and genomics is critical for precision oncology but remains challenged by the quadratic computational complexity and noise susceptibility of Transformer-based fusion mechanisms. To address these problems, we present ME-Mamba, a Multi-Expert Mamba framework that achieves linear complexity while enabling robust cross-modal interaction. Specifically, we propose a novel attention-guided scan strategy to mitigate the sequential scanning bias inherent in Mamba, ensuring the prioritization of discriminative features regardless of spatial order. Furthermore, we introduce a Synergistic Expert equipped with a parameter-free, dual-granularity fusion mechanism. By combining Optimal Transport (OT) for precise local alignment and Maximum Mean Discrepancy (MMD) for global distribution consistency, our approach acts as a hard filter to maximize the signal-to-noise ratio while preventing overfitting. Extensive experiments on five TCGA datasets demonstrate that ME-Mamba significantly outperforms state-of-the-art methods in both prediction accuracy and computational efficiency. We will make our code publicly available.