Pathology and pathway fusion drops all-to-all matching: a dynamic sparse graph plus Mamba lifts mean C-index from 0.609 to 0.638 across five TCGA cohorts
Synopsis
Multimodal survival prediction from whole-slide pathology images and genomic pathways has relied on dense or fixed-rule cross-modal fusion, which can add redundant interactions and miss patient-specific pathology-molecular links. The new method lets pathway tokens query slide patches and keeps only the Top-k highest-scoring patches for each pathway, forming a patient-specific sparse bipartite graph that graph attention then refines before Mamba state-space blocks aggregate the multimodal token sequence. Under patient-level five-fold cross-validation on five TCGA cohorts, mean C-index rose from 0.609 for SurvPath to 0.638, an absolute gain of 0.029; the paired t-test reached significance only in BLCA (p = 0.009), while other cohorts showed positive but non-significant mean differences.
Interpretation
Cross-modal fusion in multimodal survival prediction moves from dense all-to-all or fixed-rule edges to a sample-adaptive selection that keeps only a small number of pathology-pathway connections per patient. Earlier Transformer-style fusion such as SurvPath models dense pathway-histology interactions, while graph methods typically build edges with predefined rules such as k-nearest neighbors or Euclidean distance, which may not express patient-specific cross-modal relations. On five TCGA cohorts (BRCA, BLCA, HNSC, LUAD, STAD) with patient-level five-fold cross-validation and identical split files, mean C-index rose from 0.609 for SurvPath to 0.638, an absolute gain of 0.029; the paired t-test was significant only in BLCA (p = 0.009), and in HNSC TransMIL(KP) reached 0.573 versus 0.572 for this method.
The pipeline runs in three steps: encode each modality separately, let pathway tokens query slide tokens to build and refine a sparse graph, then aggregate with Mamba. Pathway tokens serve as queries and slide patches as keys and values; each pathway keeps only its Top-k patches by attention score (k = 8 in the main setting), forming a patient-specific bipartite graph whose selected pathway representations graph attention then updates. Mamba's selective state space propagates information along long sequences with input-dependent parameters instead of dense pairwise self-attention; under matched profiling, peak GPU memory was 441 MB versus 630 MB for the Transformer replacement, with 22.1 minutes per training epoch and 15.6 ms inference per patient.
Perspective
The conclusions rest on patient-level five-fold cross-validation over five TCGA cohorts (BRCA, BLCA, HNSC, LUAD, STAD), with disease-specific survival as the endpoint, UNI-derived slide features and genomic features grouped by Hallmark pathways. All comparisons use the same matched patients, identical split files and identical inputs, and baselines were re-implemented or adapted and evaluated on these cohorts rather than copied from their original publications. External validation, repeated-seed evaluation and pathologist-guided interpretation remain necessary before clinical application.
Mean differences are positive but not statistically significant outside BLCA, and five folds per cohort give limited statistical power; fold-level standard deviations range from 0.061 to 0.092, with LUAD the most variable. All experiments use TCGA, and patient-level cross-validation does not establish robustness to institutional, scanner, staining, preprocessing or molecular batch variation; main results use random seed 1 and validation-selected hyperparameters without repeated-seed evaluation. Pathway-patch visualizations have not undergone pathological or biological validation and should be read as model-derived hypotheses, and the main performance table lacks a dedicated recent graph-based WSI survival baseline.
