Integrating Multi-Level Features of Pathological Images and Pathway Features for cancer Survival Prediction

Accurate cancer survival prediction is crucial for personalized precision medicine. However, existing survival analysis methods based on Whole Slide Images typically rely on single-magnification image patch features, neglecting key multi-magnification level information in histopathological images and failing to effectively integrate biological features such as pathways. To address these limitations, this paper proposes a novel multi-instance learning model called HiMulti, which improves survival prediction accuracy by deeply fusing multi-level pathological image features and pathway features. HiMulti first constructs WSI into multiple multi-level pyramid-structured regions of interest. Then, it utilizes improved Mamba-Inspired Linear Attention model(iMILA) and linear attention Transformer to capture intra-level micro-texture details and inter-level macro-structural interactions, respectively. Simultaneously, the model incorporates pathway features based on RNA-Seq and generates multimodal and unimodal attention weights through a novel dual-branch attention mechanism to accurately select key image regions with the highest prognostic value. Finally, patient-level survival prediction is achieved by aggregating these key region features. Experiments on the TCGA-LUAD, BRCA, and BLCA datasets show that HiMulti's C-Index is significantly improved by 4.45% to 21.01% compared to state-of-the-art weakly supervised methods, and visualization results further confirm the model's superiority.