Causal Spectral Segmenter: Counterfactual Graph Reasoning for Weakly Supervised Pathology Segmentation.

Abstract

Weakly supervised pathology segmentation aims to alleviate the reliance on costly pixel-level annotations, yet remains challenging due to severe tissue heterogeneity, complex morphological patterns, and strong contextual confounding in histopathology images. Existing transformer-based methods often learn spurious correlations between lesions and surrounding tissues, while graph-based approaches tend to suffer from feature over-smoothing, resulting in degraded boundary delineation and localization accuracy. To address these challenges, we propose a novel Causal Spectral Segmenter (CSS) for weakly supervised pathology segmentation. The proposed framework seamlessly integrates causal representation learning and multi-scale spectral graph reasoning. Specifically, a Causal Counterfactual Projector (CCP) is introduced to estimate lesion-specific causal effects through factual-counterfactual intervention, thereby suppressing contextual confounding and enhancing lesion-discriminative representations. Furthermore, we develop a Multi-scale Spectral Graph Reasoner (MSGR) composed of stacked Spectral Chebyshev Graph Convolution (SCGC) layers, which perform topology-aware spectral propagation across multiple neighborhood scales to capture long-range tissue dependencies while mitigating graph over-smoothing. By jointly modeling causal effects and multi-scale topological structures, CSS effectively improves lesion localization and boundary preservation under weak supervision. Extensive experiments on two public histopathology segmentation benchmarks demonstrate that CSS consistently outperforms state-of-the-art methods and achieves superior segmentation accuracy and structural consistency. The source code and implementation details will be publicly available at: https://github.com/zhangxu90s/CSS.