Semantic-Anchored Evidential Fusion for Domain-Robust Whole-Slide Survival Analysis

Computational pathology models for cancer survival analysis often fail when deployed across different clinical centers due to domain shifts caused by variations in staining protocols and scanner hardware. This study investigates whether high-level pathology semantics—such as tumor grade and micro-environmental architecture—can serve as domain-invariant anchors to improve the robustness of survival predictions without requiring target-domain data.

The authors propose the Semantic-Anchored Evidential Fusion Survival (SAEFS) framework. Instead of relying solely on pixel-level features, SAEFS uses a pathology-specialized Visual Question Answering (VQA) model to extract structured, clinically grounded semantic descriptions from whole-slide images (WSIs). These semantic anchors guide a dual-stream evidence extraction process: one stream focuses on text-guided visual features, while the other captures holistic visual context. To integrate these potentially correlated sources, the model employs Dirichlet-based Subjective Logic and a cautious conjunction rule, which explicitly models uncertainty and prevents the overconfident fusion of evidence.

SAEFS demonstrates significant improvements in cross-center generalizability. When trained on a single source domain and evaluated zero-shot across four unseen clinical cohorts, the model consistently outperformed state-of-the-art baselines, achieving an average C-index improvement of 10.2%. Quantitative analysis confirms that VQA-derived semantic features exhibit substantially lower cross-center divergence compared to traditional pixel-derived features, validating the hypothesis that high-level biological concepts are more stable across different clinical environments than low-level visual statistics.