Causality-Guided Diffusion and Fusion of incomplete multi-modal data for robust survival prognosis

Accurate integration of whole-slide images (WSIs) and genomic data is essential for improving the reliability and interpretability of survival prognosis. However, current methods face core challenges of insufficient robustness in both cross-modal fusion and the handling of incomplete data. To reduce computational burden, current approaches typically compress WSIs independently, which disconnects the inherent biological links between histopathology images and genomic information. Meanwhile, prevalent attention-based fusion methods rely heavily on fitting statistical correlations from data, making it difficult to distinguish genuine biological associations from spurious statistical dependencies. Moreover, when genomic data is incomplete, most methods perform feature imputation solely by learning from data distributions, without ensuring the biological plausibility of the imputed features. To address these issues, we propose a Causality-Guided Diffusion and Fusion model (CGDF), which jointly regularizes both feature generation and fusion through a Causal Effect Matrix. First, we design a cross-modal prototype learning method that is based on mutual information optimization to compress WSIs features. We then construct a Causal Graph Convolutional Network to learn a Causal Effect Matrix, which guides a Diffusion Transformer in generating biologically plausible genomic features for missing data. Finally, we introduce a Causal Attention Fusion Network to achieve robust cross-modal integration. Experiments on five public TCGA datasets demonstrate that CGDF outperforms state-of-the-art methods in survival prediction and clinical grading tasks under conditions of complete data and missing genomic data.