AdaSurvMamba Enhances Multimodal Cancer Prognosis with Dynamic AI

Multimodal survival analysis, which combines data like whole slide images (WSIs) and genomic profiles, is crucial for accurate cancer prognosis. While state-space models such as Mamba have shown promise in sequence modeling, their application to complex multimodal tasks faces two key challenges. First, traditional fusion methods assume a constant interaction strength between modalities, overlooking how the diagnostic importance of each data type can fluctuate across patients and regions. Second, Mamba's standard architecture processes tokens along fixed physical paths, which can disrupt the semantic continuity of spatially dispersed medical features and weaken long-range contextual understanding. To overcome these limitations, researchers introduce AdaSurvMamba, an innovative adaptive framework for multimodal survival analysis. This framework incorporates a Dual-Scale Importance-Aware Reconstruction (DSIR) module, which dynamically adjusts the cross-modal interaction strength by assessing diagnostic importance at both sequence and token levels. Additionally, it features a Semantic Aggregation Scanning (SAS) module that reorganizes discrete tokens into semantically continuous sequences using a shared prototype pool. SAS also adaptively controls information absorption by modulating the state transition step size based on global modality context and semantic priors. Experiments conducted across five TCGA (The Cancer Genome Atlas) cohorts demonstrated that AdaSurvMamba consistently achieved superior results compared to existing methods, highlighting its potential for more accurate cancer prognosis.