Prior-guided whole-slide survival prediction with soft histopathologic priors from vision-language foundation models

Abstract Whole-slide image (WSI) survival prediction is complicated by heterogeneous tumor morphology, where prognostic evidence is distributed across tumor, stroma, immune infiltration, necrosis, and residual tissue regions and often appears as mixed patch-level tissue evidence. Existing models commonly aggregate patch embeddings directly or convert tissue cues into hard pseudo-labels, which can discard uncertainty at tumor-microenvironment interfaces. We propose a prior-guided WSI survival framework that uses soft histopathologic priors derived from the CONCH vision-language foundation model. Prompt-similarity scores are normalized into patch-level tissue distributions and used to guide probability-weighted tissue-prototype initialization, entropy-weighted prior-consistency learning, and prototype-conditioned graph-context aggregation over a spatial k-nearest-neighbor graph. The residual Others class is retained for representation learning but excluded from prognostic prior alignment. In patient-level 5-fold cross-validation on TCGA KIRC, KIRP, LUAD, STAD, and UCEC cohorts, the model achieved C-index values of 0.7307, 0.8153, 0.6375, 0.6744, and 0.7029, respectively, matching or exceeding the strongest baseline in every cohort. Kaplan-Meier analysis showed significant high-versus low-risk separation across all cohorts, and ablation studies confirmed the contributions of soft priors, entropy weighting, the Others channel, and graph-context aggregation. These results show that retaining patch-level histopathologic uncertainty provides a consistent and interpretable signal for WSI-based survival risk stratification.