Search papers, labs, and topics across Lattice.
This paper introduces C$^2$RM-Seg, a two-stage framework for weakly supervised histopathological tissue segmentation that addresses the limitations of existing Class Activation Mapping (CAM) methods by integrating causal pseudo-label refinement with structural-semantic enhancement. The Causal Counterfactual Reasoning Module (C$^2$RM) decomposes features into latent factors and applies counterfactual interventions to generate morphology-aligned CAMs, while the Dual-Path Structural-Semantic Architecture combines structural features with semantic priors to improve segmentation accuracy. Experimental results demonstrate that C$^2$RM-Seg outperforms existing methods on two public datasets, achieving state-of-the-art performance in tissue segmentation tasks.
C$^2$RM-Seg achieves state-of-the-art histopathological tissue segmentation by effectively aligning causal reasoning with structural semantics, overcoming the noise of traditional CAM methods.
Histopathological tissue segmentation is essential for computer-aided diagnosis, yet weakly supervised methods often suffer from noisy pseudo-labels generated by Class Activation Mapping (CAM). Existing CAM approaches tend to focus on staining-driven appearance cues rather than true causal tissue morphology, resulting in spurious localization and poor structural consistency. To address this issue, we propose C$^2$RM-Seg, a two-stage framework that integrates causal pseudo-label refinement with structure-aware semantic enhancement. For classification, we introduce a Causal Counterfactual Reasoning Module (C$^2$RM) that decomposes features into latent factors and performs counterfactual intervention via a learned causal structure matrix, suppressing confounding context and producing morphology-aligned CAMs. For segmentation, we design a Dual-Path Structural-Semantic Architecture that combines fine-grained structural features from ResNeSt with global semantic priors from a frozen DINOV3 foundation model. A cross-path gating mechanism adaptively regulates semantic injection using local structural cues to preserve boundary fidelity. To further mitigate residual pseudo-label noise, we propose an Uncertainty-Gated Margin (UGM) loss, which dynamically balances margin enforcement and confidence learning based on prediction uncertainty. Extensive experiments on two public histopathological tissue datasets show that C$^2$RM-Seg achieves state-of-the-art performance.
