Source-linked AI summary

Synergistic Information Disentanglement for Omni-modal Slide Representation Learning in Computational Pathology

Mingxin Liu, Chengfei Cai, Anwen Lu, Pengbo Xu, Jun Li, Jinze Li, Depin Chen, Jun Xu

arXiv:2609.02118v1cs.CV

TL;DR

Multimodal CPath representation learning can under-use complementary information by aligning modalities around shared redundancy, overlooking synergistic diagnostic signals. Φ-Omni addresses this with an SIB module and ΦID objective grounded in PID theory, and reports superior few-shot performance across independent datasets and tasks against supervised and SSL baselines.

  • Problem

    Existing multimodal CPath methods rely on contrastive alignment that can overlook synergistic information absent from individual modalities, while pathology reports remain under-utilized.

  • Method

    Φ-Omni uses domain-specific encoders, an SIB module, and a ΦID objective to distill high-order cross-modal interactions while minimizing shared redundancy.

  • Results

    Φ-Omni achieves superior few-shot performance across eight tasks and five independent cohorts compared with supervised MIL and SSL baselines.

  • Takeaways & Limitations

    The framework supports slide representations that transfer to unimodal few-shot inference while retaining synergistic information from omni-modal pretraining.

Abstract

from arXiv · show

In computational pathology (CPath), developing omni-modal self-supervised learning (SSL) models that integrate histology, genomics, and clinical reports enables transferable representation learning for whole slide images (WSIs). Existing approaches implicitly force heterogeneous modalities into a uniform latent space by contrastive alignment, causing modality collapse where unique, synergistic diagnostic signals (termed as $\mathrmΦ$) are discarded in favor of trivial redundancy. We hypothesize that the strongest task-agnostic SSL training signal stems from distilling the synergistic interactions over merely aligning shared redundancy. To this end, we introduce \textsc{$\mathrmΦ$-Omni}, a synergistic information disentanglement framework grounded in Partial Information Decomposition (PID) theory for slide representation learning. Unlike standard contrastive approaches, \textsc{$\mathrmΦ$-Omni} employs a Synergistic Information Bottleneck (SIB) regulated by the proposed $\mathrmΦ\text{ID}$ objective, which explicitly suppresses marginal redundancy while maximizing irreducible synergy, thereby distilling high-order cross-modal interactions. Following pretraining on breast ($n$=1031) and lung ($n$=919) cohorts, \textsc{$\mathrmΦ$-Omni} demonstrates superior few-shot performance across five independent external datasets spanning eight tasks compared to supervised and SSL baselines. Source code is available here.

1 Introduction

Φ-Omni addresses limitations of multimodal slide representation learning by disentangling synergistic cross-modal information rather than relying only on redundancy-driven alignment. Its SIB and ΦID components support compact omni-modal pretraining and unimodal few-shot inference.

  • Background: The framework builds on a CPath workflow in which WSIs are tessellated, patch features are extracted, and embeddings are aggregated for transferable slide representations.SSL slide embeddings can support zero- or few-shot transfer without task-specific fine-tuning.
  • Motivation: Contrastive alignment can overlook synergistic diagnostic information that emerges only from jointly observing pathology, genomics, and clinical reports.Existing methods may under-utilize pathology reports and favor shared redundancy, despite diagnosis relying on complementary cross-modal cues.
  • Method: Φ-Omni encodes omni-modal data with domain-specific encoders and integrates the embeddings through a Synergistic Information Bottleneck.The SIB module is designed to distill high-order cross-modal interactions into the slide representation.
  • Method: The ΦID objective maximizes synergistic interactions while minimizing shared redundancy in the omni-modal latent space.This replaces redundancy alignment with synergistic disentanglement as the central SSL objective.
  • Evaluation: Φ-Omni validates superior few-shot generalizability across eight diagnostic tasks and five public independent cohorts against supervised MIL and SSL baselines.The frozen slide encoder supports unimodal few-shot adaptation at inference.

2 Method

Φ-Omni encodes pathology, genomics, and report data with domain-specific encoders, then uses SIB and ΦID to suppress redundancy while distilling synergistic cross-modal information. Its training combines symmetric contrastive alignment with a ΦID regularizer that compresses marginal information and expands joint information.

  • Omni-modal Feature Encoding: Domain-specific encoders project pathology slides, genomic profiles, and text reports into modality-specific embeddings for omni-modal learning.The pathology pipeline uses WSI patch embeddings and multiple-instance aggregation; genomics and reports are separately projected into embeddings.
  • Synergistic Information Bottleneck: SIB maps concatenated unimodal embeddings from 3d to d through a 3d → 2d → d bottleneck, filtering modality-specific noise while retaining synergistic diagnostic signals.The bottleneck is implemented with projection layers and Layer Normalization.
  • Synergistic Information Disentanglement: PID decomposes mutual information into redundancy, unique modality information, and synergy, motivating Φ-Omni’s focus on information available only through joint multimodal observation.The method treats standard mutual-information maximization as insufficient because it can favor shared redundancy and modality collapse.
  • Synergistic Information Disentanglement: ΦID maximizes the whole-minus-sum synergistic residual while minimizing marginal redundancy inferred from individual modalities.The objective uses projected marginal and joint representations to separate complementary cross-modal interactions from unimodal information.
  • Gaussian Canonical Projector: GCP maps heterogeneous embeddings into a smooth Gaussian latent space, enabling tractable mutual-information estimation through differential-entropy approximations.The joint covariance log-determinant represents feature volume, which is expanded to prevent representation collapse; ε = 10^-5 provides numerical stability.
  • Pretraining Objective: Pretraining combines symmetric pathology-centered InfoNCE alignment with ΦID regularization as L = LSymCL + γLΦID, using γ = 0.2.The contrastive term aligns pathology with genomics and reports, while ΦID penalizes trivial redundancy and encourages complementary features.

3 Experimental Setups and Results

Φ-Omni is evaluated through omni-modal pretraining, few-shot transfer across independent cohorts, and ablations of architecture, fusion, and loss design. Results report broad gains over MIL and SSL baselines, with synergy disentanglement improving performance across six tasks.

  • Study Design: Pretraining uses strictly triplet-aligned pathology WSIs, genomic profiles, and pathology reports, while inference requires only WSIs.Breast and lung cohorts contain 1,031 and 919 cases, respectively.
  • Few-shot Results: Φ-Omni surpasses all MIL baselines in 8/8 tasks, including gains of +4.3%, +1.2%, and +1.1% over ABMIL for molecular prediction, breast subtyping, and lung subtyping.Φ-Omni uses linear probing, whereas MIL models are trained from scratch in the few-shot setting.
  • Few-shot Results: Φ-Omni outperforms SSL models on most tasks, including +3.2% over TANGLE on BRACS and +5.7% over UNIv2 on CPTAC-LSCC.The advantage is reported consistently across all k values and tasks.
  • Ablation Study: SIB outperforms Trilinear and Concat fusion across all six tasks, while the selected ABMIL pre-attention architecture improves over MLP, TransMIL, and ABMIL.These comparisons constitute the architecture ablation.
  • Ablation Study: +4.4% on BRACS-Subtyping and +7.2% on BRCA-PIK3CA result from combining SymCL with ΦID rather than using SymCL alone.The ablation identifies ΦID as indispensable for the reported synergistic improvements.
  • Interpretability Analysis: Attention heatmaps align high-attention regions with tumor areas, while t-SNE shows three separated modality clusters with preserved cross-modal connections.The analysis covers lung and breast WSIs and pathology, transcriptomics, and text-report embeddings.

4 Conclusion

Φ-Omni shifts omni-modal slide representation learning from redundancy-driven contrastive alignment toward synergistic information disentanglement. Its interpretability analysis reports tumor-focused attention and separated modality clusters, while the paper concludes that synergy distillation supports data-efficient multimodal pathology frameworks.

  • Conclusion: Φ-Omni uses PID-derived SIB and ΦID mechanisms to extract high-order diagnostic insights absent from individual modalities.The conclusion frames this as a shift from redundancy-driven alignment to synergistic information disentanglement.
  • Conclusion: The paper reports state-of-the-art few-shot performance across six benchmarks together with biologically grounded interpretability.The conclusion connects these outcomes to omni-modal histology, genomics, and pathology reports.
  • Conclusion: The authors position synergistic interaction distillation as a basis for data-efficient and robust multimodal computational pathology frameworks.This statement is presented as the work's broader supported implication.
  • Interpretability: Attention heatmaps show high-attention regions corresponding to tumor areas in randomly selected lung and breast WSIs.Red and blue indicate high and low attention, respectively.
  • Interpretability: t-SNE visualization separates pathology, transcriptomics, and text-report embeddings into distinct clusters while retaining cross-modal connections.The visualization is presented as evidence against modality collapse and for preserved biological and clinical correlations.
Loading 2609.02118v1…