PaperlayerView exact revision on arXiv

Source-linked AI summary

Modeling Whole-Slide Images as Dynamic Tumor Microenvironment Fields

Lei Wu, Jiashuai Liu, Di Zhang, Zhangpeng Gong, Yingkang Zhan, Yi Niu, Jiusong Ge, Chunze Yang, Kai Yi, Mireia Crispin-Ortuzar, Chen Li, Zeyu Gao

arXiv:2609.34451v1cs.CVcs.AI

TL;DR

Weakly supervised WSI models often lack explicit tissue-level organization and microenvironment interaction modeling. TMEvolve forms adaptive regions and evolves them through diffusion and concept-guided boundary flux, reporting strong performance across six cohorts and three tasks.

  • Problem

    Many WSI models lack an explicit mechanism to organize patches into tissue-level structures and model interactions among microenvironment regions.

    Source: §1 · pp. 1–2

  • Method

    TMEvolve forms adaptive soft tissue regions and evolves them over patch neighborhoods through intra-region diffusion and concept-guided boundary flux.

    Source: §1 · p. 2

  • Results

    Across six cohorts, TMEvolve achieves strong performance against representative MIL methods, pathology foundation models, and survival-oriented baselines across survival prediction, gene expression prediction, and subtype classification.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Ablation and visualization analyses support dynamic microenvironment modeling as a promising direction for weakly supervised computational pathology.

    Source: §5 · p. 10

    Pseudo-time evolution represents iterative feature refinement, not chronological tumor-microenvironment dynamics, and the concept bank may have limited coverage of rare or dataset-specific patterns.

    Source: G Limitations · p. 24

Abstract

from arXiv · show

Due to the gigapixel-scale nature of whole-slide images (WSIs), weakly supervised WSI analysis is commonly formulated as a multiple instance learning (MIL) problem, where patch-level features are aggregated into slide-level representations. However, diagnostic and prognostic evidence often arises from spatially coherent tumor microenvironment regions and their interactions, rather than isolated patches alone. Existing patch-level or static region-based methods usually overlook how tissue regions should be adaptively formed and subsequently evolved through microenvironment interactions across heterogeneous boundaries. In this paper, we propose Concept-Guided Tumor Microenvironment Evolution (TMEvolve), a reaction-diffusion-inspired framework that models WSIs as latent tumor microenvironment fields over discrete patch graphs. TMEvolve instantiates this view as a learnable graph-discretized evolution process over patch neighborhoods. It first forms adaptive soft tissue regions as coherent microenvironment units, then performs pseudo-time evolution through two complementary local dynamics: intra-region diffusion, which stabilizes latent states within coherent tissue compartments, and concept-guided boundary flux, which propagates visual feature signals and language-derived concept signals across heterogeneous region interfaces. The evolved microenvironment regions are finally aggregated for slide-level prediction. We evaluate TMEvolve on six datasets across three weakly supervised WSI tasks: survival prediction, gene expression prediction, and histological subtype classification. TMEvolve consistently improves over representative MIL methods, pathology foundation models, and concept-guided baselines. Ablation studies and visualizations further support the effectiveness and interpretability of TMEvolve, highlighting the value of dynamic region modeling and boundary interaction.

1 Introduction

Existing weakly supervised WSI methods often lack explicit tissue-level organization and modeling of interactions among microenvironment regions. TMEvolve forms adaptive regions, evolves them through local diffusion and concept-guided boundary flux, and reports strong performance across three tasks.

1 Introduction

  • Most WSI-level models lack an explicit mechanism to organize patches into tissue-level structures and model interactions among microenvironment regions.Patch-level MIL overlooks region organization, while fixed grids, heuristic partitions, or static neighborhoods may cross natural boundaries and mix heterogeneous components.Source: §1 · pp. 1–2
  • Coherent neighboring patches motivate local diffusion, while heterogeneous interfaces motivate concept-guided modeling of cross-region flux and reaction.The paper frames these distinct local conditions as a reaction–diffusion-inspired dynamic view of WSI representation learning.Source: §1 · p. 2
  • TMEvolve forms adaptive soft tissue regions and evolves them through intra-region diffusion and concept-guided boundary flux before aggregating regions for slide-level prediction.The framework uses learnable graph-discretized latent evolution over patch neighborhoods rather than solving a reaction–diffusion PDE.Source: §1 · pp. 2–3
  • Across six cohorts, TMEvolve achieves strong performance against representative MIL methods, pathology foundation models, and survival-oriented baselines across three WSI tasks.Ablation and visualization results support its effectiveness and interpretability.Source: §1 · pp. 2–3

2 Related Work

Prior WSI methods aggregate patches, use pretrained or language-guided representations, or incorporate fixed spatial context. TMEvolve instead models concept-bearing tissue states and their interactions during pseudo-time evolution.

Weakly supervised MIL baselines for WSI prediction

  • Representative MIL methods differ mainly in patch-to-slide aggregation and generally do not explicitly model interactions among heterogeneous tissue states.Examples include ABMIL [23], DSMIL [28], TransMIL [38], and RRT-MIL [39].Source: §2 · p. 3

Pathology foundation and vision-language models

  • Pathology foundation and vision-language methods use pretrained representations or semantic guidance, while TMEvolve uses concepts to guide interactions among tissue states during pseudo-time evolution.The related methods include image-language alignment and textual-concept approaches such as PLIP, Quilt [22], CONCH, CITE [48], QPMIL [20], ConcepPath [49], and GECKO [27].Source: §2 · p. 3

Spatial and microenvironment-aware WSI modeling

  • Spatially aware models incorporate local context or relationships, but often use spatial structure as fixed context rather than dynamically refining tissue states and boundary interactions.TMEvolve is motivated by microenvironment organization and interaction, including immune infiltration, tumor invasion, immune exclusion, and stromal remodeling.Source: §2 · p. 3

3 Method

TMEvolve models a whole-slide image as an adaptive tumor microenvironment field and evolves its patch-graph representations through region-level diffusion and concept-guided boundary interactions. Evolved region states are aggregated for slide-level prediction.

3.1 WSI as a Latent Reaction–Diffusion Microenvironment Field

  • TMEvolve formulates WSI representation learning as graph-discretized latent reaction–diffusion, with within-region stabilization, anisotropic boundary interactions, and concept-conditioned semantics.Source: §3 · p. 3, §3.1 · p. 4

3.2 Adaptive Region Field and Within-region Diffusion

  • Soft patch-to-region assignments form morphologically consistent, spatially continuous compartments, while membership-overlap weights enable diffusion within regions and suppress cross-boundary smoothing.Source: §3.2 · pp. 4–5

3.3 Concept-guided Boundary Reaction–Flux Evolution

  • Boundary evolution couples feature-level anisotropic flux with concept-conditioned semantic reaction, using region heterogeneity, concept-response contrast, and directional compatibility to regulate interfaces.Source: §3.3 · pp. 5–6

3.4 Evolution-aware Aggregation and Task Prediction

  • After each pseudo-time step, patch states refresh the soft region assignments and prototypes; evolved region states are then attention-aggregated for task-specific slide prediction.Source: §3.3 · p. 6, §3.4 · p. 6

4 Experiments

Experiments evaluate TMEvolve across six datasets and three WSI tasks, comparing it with established methods and examining its components, hyperparameters, and visual interpretability. Results report strong comparative performance, benefits from the full design, and task-dependent sensitivity to model settings.

4.1 Experimental Setup

  • Six datasets cover survival prediction, gene expression prediction, and fine- and coarse-grained subtype classification, using C-index, Pearson correlation, AUC, and accuracy.Survival uses three TCGA cohorts and a private GBC cohort; gene prediction uses LUAD, BLCA, and BRCA, while subtype tasks use BRACS and EBRAINS.Source: §4.1 · p. 7

4.2 Comparison with State-of-the-art Methods

  • TMEvolve achieves the best survival performance on three of four cohorts and strong or competitive subtype and gene-prediction performance against representative methods.The comparison spans MIL, pathology foundation or WSI-level, task-oriented, and concept-guided baselines; Figure 3 reports gene-level Pearson correlation across TCGA tasks.Source: §4.2 · pp. 7–8, Figure 3 · p. 8

4.3 Ablation Study and Hyperparameter Sensitivity

  • Removing intra-region diffusion, pseudo-time evolution, or diversity loss degrades performance; random concept embeddings also weaken results, especially for gene expression prediction.The full model performs best across tasks, supporting the joint effectiveness of region stabilization, concept guidance, and pseudo-time evolution.Source: §4.3 · p. 8, Table 3 · p. 8
  • Subtype classification works well with fewer evolution steps and coarser regions, while survival and gene prediction benefit from more refined modeling.Moderate settings such as Q = 3 and K = 6 or K = 8 generally perform better; further increases bring no consistent gain and may add redundant interactions or over-smoothing.Source: §4.3 · p. 9, Table 4 · p. 9

4.4 Interpretability Analysis

  • In representative LUAD cases, pseudo-time evolution consolidates patch assignments into coherent, concept-aligned regions, with boundary maps localizing heterogeneous interfaces.High-risk cases show more prominent adverse tumor-related concept responses, while low-risk cases show relatively stronger immune- or glandular-related patterns.Source: §4.4 · p. 9, Figure 4 · p. 9

5 Conclusion

TMEvolve models WSIs as dynamic tumor microenvironment fields, evolving adaptive concept-bearing regions across heterogeneous tissue boundaries. Experiments and analyses support its effectiveness and interpretability across three WSI tasks.

5 Conclusion

  • TMEvolve models WSIs as dynamic tumor microenvironment fields and evolves adaptive concept-bearing regions across heterogeneous boundaries, with experiments supporting effectiveness across three tasks.Ablation and visualization analyses support intra-region diffusion, concept guidance, pseudo-time evolution, and interpretable boundary-level evidence.Source: §5 · p. 10

Supplementary Material

The supplement provides additional method details, extended experiments, and discussions for TMEvolve.

Supplementary Material

Contents of the Supplement

The supplement derives TMEvolve’s latent reaction–diffusion formulation as a learnable evolution process on a patch-neighborhood graph. Its update combines within-region diffusion, boundary-localized flux, and a concept-conditioned semantic source.

A Graph Discretization of the Latent Reaction–Diffusion Formulation

  • The derivation represents patch locations as nodes in a spatial neighborhood graph and patch features as the latent states.Source: A Graph Discretization of the Latent… · p. 15
  • Within-region diffusion uses region-membership overlap as graph conductivity, while boundary evolution uses anisotropic conductivity activated by heterogeneity, concept-response contrast, and directional compatibility.Source: A Graph Discretization of the Latent… · p. 15
  • The update’s three summations encode within-region diffusion, boundary-localized anisotropic flux, and a concept-conditioned semantic reaction source.Conductivities and concept assignments are recomputed at each pseudo-time step, making evolution state-dependent rather than fixed graph smoothing.Source: A Graph Discretization of the Latent… · p. 15

B.1 Dataset Details

The experiments cover three weakly supervised WSI tasks across six cohorts or datasets. They use five-fold cross-validation, with task-specific region counts and evolution steps.

B.1 Dataset Details

  • Experiments span survival prediction, gene expression prediction, and subtype classification across TCGA cohorts, a private GBC cohort, BRACS, and EBRAINS.The subtype tasks use fine- and coarse-grained label settings; gene expression prediction covers 25 genes per cohort.Source: §B.1 · p. 15

B.2 Hyperparameter Settings

  • All experiments use five-fold cross-validation, with standard K-fold splits for survival and gene expression prediction and stratified K-fold splits for subtype classification.Source: §B.2 · p. 16
  • Task- and dataset-specific region counts K and evolution steps Q are summarized alongside the main implementation settings.Source: Table 7 · p. 16, Table 6 · p. 16, §B.2 · p. 16

C Computational Cost Analysis

The cost analysis reports TMEvolve’s parameter count and per-slide runtime, while the concept-construction section describes shared and dataset-specific semantic priors.

C Computational Cost Analysis

  • 5.18M trainable parameters and 6.48 ms per slide are reported for TMEvolve on LUAD survival prediction.The comparison uses pre-extracted patch features, batch size 1, and excludes offline feature extraction.Source: C Computational Cost Analysis · p. 16, Table 8 · p. 17
  • The additional cost comes mainly from dynamic region assignment, local coordination, and concept-guided boundary interaction, with runtime remaining in the millisecond range.The paper characterizes this as a trade-off for explicit microenvironment modeling capacity.Source: C Computational Cost Analysis · p. 17

D Pathological Concept Construction

  • Language-encoded pathological concepts provide semantic priors for region routing and concept-guided boundary interaction.The descriptions cover tissue states associated with tumor morphology, immune activity, stromal response, invasion, necrosis, vascular remodeling, and organization.Source: D Pathological Concept Construction · p. 17
  • TMEvolve combines a shared concept bank of broadly applicable patterns with dataset-specific concepts capturing cohort-dependent characteristics.The random-concept ablation substitutes same-dimensional random probes for language-encoded embeddings, removing the semantic prior while keeping parameter budget comparable.Source: D Pathological Concept Construction · p. 17

E Additional Gene Expression Prediction Results

The appendix extends gene-expression evaluation, survival-stratification evidence, and interpretability visualizations. Across the survival folds, predicted risk groups are generally separated, and visualizations show region refinement and concept-related patterns.

E Additional Gene Expression Prediction Results

  • Full gene-level evaluation covers TCGA-LUAD, BLCA, and BRCA using Pearson and Spearman correlations and AUC after median-based expression binarization.The representative genes were selected for clinical relevance and biological interpretability rather than model performance.Source: E Additional Gene Expression… · pp. 17–18

F.1 Five-fold Kaplan–Meier Survival Curves

  • TMEvolve generally separates high- and low-risk groups across cohorts in the five cross-validation folds.Figure 5 presents fold-specific Kaplan–Meier curves for BLCA, LUAD, GBC, and BRCA.Source: §F.1 · p. 23, Figure 5 · p. 23

F.2 Additional Interpretability Visualization

  • Representative WSI visualizations show region partitions becoming more spatially coherent after evolution, with boundary and concept maps highlighting heterogeneous interfaces and concept-related patterns.Source: Figure 6 · p. 24, §F.2 · p. 24

G Limitations

The limitations distinguish TMEvolve’s representation-learning role from biological simulation and identify constraints in learned regions, concept coverage, and clinical use.

G Limitations

  • Pseudo-time evolution is iterative feature refinement, not a mechanistic simulation of true tumor-microenvironment progression.Source: G Limitations · p. 24
  • Expert-reviewed concept banks may still have limited coverage of rare or dataset-specific patterns.Source: G Limitations · p. 24
  • Region partitions learned from slide-level supervision without pixel-level tissue annotations are model-derived organizations, not exact pathological segmentation masks.Source: G Limitations · p. 24

H Broader Impact

  • Clinical deployment requires rigorous external validation and expert oversight; the paper positions TMEvolve as research and decision support, not a standalone diagnostic system.The broader-impact discussion also flags dataset bias, domain shifts, and over-interpretation of model-derived regions as risks without sufficient validation.Source: H Broader Impact · p. 25

Related papers

Loading 2609.34451v1…