Enhancing Interpretation of Histopathology Whole Slide Image Analysis via Regional Causal Dependency Discovery

Zixian Li, Jun Shi, Zhiguo Jiang, Fengying Xie, Yushan Zheng

Keywords: Computational pathology, WSI analysis, Causal discovery.

Abstract: Histopathology whole slide image (WSI) analysis is fundamental to computational pathology. Attention-based heatmaps are commonly used for interpretability in WSI analysis. However, heatmap is limited in describing the potential relationships between multiple high-probability regions, which restricts its application in fine-grained WSI analysis tasks. In this paper, we propose Pathology Causal Discovery Network (PCDN), a novel framework that reconstructs interpretable diagnostic pathways by dynamically discovering regional causal dependencies from WSIs. Unlike approaches relying on predefined medical priors, PCDN introduces a Causal Structure Learner (CSL) to infer a Directed Acyclic Graph (DAG) which represents the causal dependencies among pathological regions. A Causal Graph Propagator (CGP) is then designed to guide feature propagation based on the DAG, integrating local causal dependencies with global context. Extensive experiments on three large-scale pathological datasets demonstrate that PCDN achieves state-of-the-art performance and can provide meaningful causal insights for WSI analysis.

Submission Number: 17