Positive Semi-definite Latent Factor Grouping-Boosted Cluster-reasoning Instance Disentangled Learning for WSI Representation

TL;DR

This paper tackles the entanglement challenges in MIL-based whole-slide image (WSI) analysis by introducing PG-CIDL, a three-phase disentangled framework. It first uses positive semidefinite latent factor grouping (PSD-LFG) to map spatially entangled patches into a latent subspace, then applies cluster-reasoning instance disentangling (CID) via counterfactual probability inference to separate semantic factors (tumor, microenvironment, background), and finally uses instance effect re-weighting to curb decision entanglement. The approach is grounded in two decoupled structural causal models and an information-theoretic weighting scheme, enabling end-to-end optimization. Empirical results on multicenter datasets show state-of-the-art accuracy and AUC, with visualizations confirming pathologist-aligned interpretability. Overall, PG-CIDL advances interpretable, causally informed WSI representations and demonstrates strong potential for clinical deployment.

Abstract

Multiple instance learning (MIL) has been widely used for representing whole-slide pathology images. However, spatial, semantic, and decision entanglements among instances limit its representation and interpretability. To address these challenges, we propose a latent factor grouping-boosted cluster-reasoning instance disentangled learning framework for whole-slide image (WSI) interpretable representation in three phases. First, we introduce a novel positive semi-definite latent factor grouping that maps instances into a latent subspace, effectively mitigating spatial entanglement in MIL. To alleviate semantic entanglement, we employs instance probability counterfactual inference and optimization via cluster-reasoning instance disentangling. Finally, we employ a generalized linear weighted decision via instance effect re-weighting to address decision entanglement. Extensive experiments on multicentre datasets demonstrate that our model outperforms all state-of-the-art models. Moreover, it attains pathologist-aligned interpretability through disentangled representations and a transparent decision-making process.

Figures (10)

• Figure 1: Motivation of PG-CIDL. (a): Conventional MIL framework with entangled representation has weak interpretbility. (b): Proposed PG-CIDL framework with spatial, semantic and decision disentanglement has better interpretability.
• Figure 2: Overview of PG-CIDL. Entangled Features extracted by patch encoder are categorized into three groups via Positive semi-definite latent factor Grouping. They are identified into tumor, microenvironment and background factors by Cluster-reasoning Instance Disentangling. End-to-end optimization is performed across all phases for disentangled representation Learning with instance effect re-weighted features.
• Figure 3: Two decoupled SCMs for disentangling factors in DRL, where node T is for factor tumor, node E for microenvironment, node G for pathological grading outcome and node $\epsilon$ for possible background noise.
• Figure 4: Performance comparison on multicentre datasets with PLIP pretrained features.
• Figure 5: ROC plots' comparison on multicentre datasets.