Title: DPsurv: Leveraging Dual-Prototype Evidential Fusion for Interpretable and Uncertainty-Aware Survival Prediction in Whole-Slide Images

Abstract:

Whole-slide images (WSIs) from pathology are extensively utilized in cancer survival analysis due to their ability to provide comprehensive histopathological details at both cellular and tissue scales. This depth of information facilitates the extraction of prognostically significant tumor features through quantitative, large-scale analysis. Despite these advantages, current WSI survival analysis techniques frequently suffer from poor interpretability and tend to neglect predictive uncertainty, particularly when dealing with the heterogeneity inherent in slide images.

To address these challenges, we introduce DPsurv, a novel evidential fusion network based on dual prototypes for whole-slide images. This framework is designed to generate survival intervals that account for uncertainty while offering clear interpretability via patch prototype assignment maps, component prototypes, and the aggregation of relative risk on a component-wise basis.

Our evaluation across five publicly accessible datasets demonstrates that DPsurv outperforms existing methods, achieving the lowest mean integrated Brier score and the highest mean concordance index. These results confirm the method's reliability and effectiveness. Furthermore, the interpretability features embedded in the prediction outcomes provide transparency across feature, reasoning, and decision-making levels, significantly boosting the trustworthiness and interpretability of the DPsurv model.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC