Topology-aware cross-modal learning with dynamic gradient modulation for survival prediction

Reliable survival prediction is critical for advancing precision oncology, yet remains challenging due to the weakly-supervised nature of whole-slide image (WSI) analysis and the inherent heterogeneity between histopathological and genomic data. Traditional methods that treat gigapixel WSIs as bags of isolated patches struggle to capture complex spatial dependencies and fall short in integrating multimodal information effectively. These limitations collectively undermine prediction robustness: spatial neglect and semantic misalignment together induce optimization imbalance during multimodal training, where one modality often dominates, thereby suppressing informative signals from the other. To address these interrelated challenges, we propose a Pathology-Genomics Balanced Fusion (PGBF) framework. A Topology-Aware Morphological Encoding (TAME) module dynamically constructs graph structures to model high-order spatial relationships among tissue regions. A Cross-Modal Semantic Alignment (CMSA) module enhances feature consistency and prioritizes diagnostically relevant regions through bidirectional alignment. Furthermore, a Gradient-Aware Dynamic Modulation (GADM) mechanism is embedded to adaptively regulate gradient flow, mitigating optimization imbalances caused by modality heterogeneity. Evaluated on five TCGA cancer cohorts, PGBF achieves an average C-index of 0.749, outperforming state-of-the-art methods. This framework that integrates GNN-based multimodal fusion with gradient balancing for survival analysis, while its modular design ensures compatibility with various attention-based or graph-driven architectures.