Conserved Immune Topology Improves Pathology Foundation Model Generalization for Cross-Cancer MSI-H Prediction

AI summary

Overview

• Research area: Computational pathology / medical computer vision — using whole-slide histopathology images and pathology foundation models to predict molecular biomarkers (microsatellite instability-high, MSI-H) across cancer types.
• Technical level: Advanced (assumes familiarity with multiple instance learning, foundation-model embeddings, and spatial statistics; the core idea is nevertheless explainable in plain terms).
• Scope: The paper proposes and evaluates Conserved Immune Topology (CIT), a ten-dimensional spatial descriptor set that augments frozen pathology foundation-model embeddings to improve zero-shot cross-cancer MSI-H prediction, evaluated on TCGA-COAD, CPTAC-COAD, and TCGA-STAD.

What This Paper Is About

Pathology foundation models combined with multiple instance learning (MIL) work well when trained and tested on the same cancer type, but they transfer poorly across cancer types because organs have different histology and architecture. The authors' goal is to give these models a representation of immune organization that is shared across organs, so that a model trained on colorectal cancer can predict MSI-H status in gastric cancer without retraining. The proposed CIT augments foundation-model tile embeddings with spatial immune descriptors computed from unsupervised clustering of tiles and their coordinates, requiring no annotations and no target-domain data.

Key Contributions

1. A lightweight spatial immune representation (CIT). CIT computes a ten-dimensional descriptor vector per tile capturing four biologically motivated groups: tertiary lymphoid structures (TLS), peritumoral margin reactions, multi-scale tumor-infiltrating lymphocyte (TIL) density, and immune–tumor mixing. It is concatenated with the frozen foundation-model embedding to form the MIL input, adding no trainable integration parameters.
2. Annotation-free and target-domain-free operation. Immune-associated tiles are identified by unsupervised K-means clustering on frozen foundation-model embeddings, using cluster tightness as an immune enrichment proxy, so no cell-level or structure-level labels and no target-domain adaptation are needed.
3. Zero-shot cross-cancer evaluation on independent cohorts. The paper reports cross-site (TCGA-COAD to CPTAC-COAD) and cross-cancer (TCGA-COAD to TCGA-STAD, and reverse-direction TCGA-STAD to TCGA-COAD) results across three MIL aggregators (ABMIL, CLAM-SB, TransMIL), with paired DeLong tests and Bonferroni adjustment.
4. Ablation and bidirectional consistency analysis. A leave-one-group-out ablation of the four descriptor families shows all groups contribute to cross-cancer transfer, with immune–tumor mixing and peritumoral margin descriptors producing the largest individual drops.

Main Findings

• Primary cross-cancer result: In zero-shot transfer from TCGA-COAD to TCGA-STAD, CIT increased TransMIL AUC from 0.6627 to 0.7161, an absolute gain of 0.0534 (98.33% DeLong CI +0.0111 to +0.0957, p = 0.003).
• All three aggregators improved on TCGA-STAD: ABMIL 0.5681 to 0.6313 (+0.0632, CI +0.0280 to +0.0984, p < 0.001); CLAM-SB 0.6242 to 0.6414 (+0.0172, CI −0.0025 to +0.0369, p = 0.037); TransMIL 0.6627 to 0.7161 (+0.0534, p = 0.003). After Bonferroni correction (α = 0.0167), only ABMIL and TransMIL improvements remained statistically significant; the CLAM-SB confidence interval spanned zero.
• Baseline cross-cancer range: Baseline models on TCGA-STAD ranged from 0.5681 to 0.6627 AUC; with CIT the range shifted to 0.6313 to 0.7161.
• Cross-site robustness (CPTAC-COAD): CIT increased CLAM-SB from 0.7989 to 0.8344 and TransMIL from 0.7870 to 0.8277. ABMIL was essentially unchanged (0.8089 vs. 0.8086, a −0.0003 difference). Cross-site gains ranged from −0.0003 to +0.0407 AUC.
• Within-cohort ceiling effect: On internal TCGA-COAD, CIT changed ABMIL by +0.001 (0.9337 ± 0.0166 to 0.9348 ± 0.0197), CLAM-SB by −0.002 (0.9469 ± 0.0233 to 0.9447 ± 0.0205), and TransMIL by +0.017 (0.9398 ± 0.0274 to 0.9567 ± 0.0112). The paper attributes this small internal effect to a ceiling effect, and notes the changes fall within fold-wise variation (SD ±0.01 to 0.06).
• Foundation model comparison: UNI2-h outperformed CONCH across all aggregators internally, with a mean Δ AUC of +0.02 to +0.03 and lower fold-wise variance (±0.02 vs. ±0.05). CONCH results were 0.9145 ± 0.0499 (ABMIL), 0.9181 ± 0.0563 (CLAM-SB), 0.9198 ± 0.0615 (TransMIL). UNI2-h was used for all cross-site and cross-cancer experiments.
• Reverse-direction transfer (TCGA-STAD to TCGA-COAD): CIT improved all three aggregators — TransMIL 0.7470 to 0.7878 (+0.041, reported as +0.0408 elsewhere in the text), ABMIL 0.7066 to 0.7228 (+0.016), CLAM-SB 0.7490 to 0.7547 (+0.006).
• Ablation ranking: Removing immune–tumor mixing (G4) reduced zero-shot TCGA-STAD AUC by 0.0111, peritumoral margin (G2) by 0.0076, TLS (G1) by 0.0046, and multi-scale TIL density (G3) by 0.0027. The two interface-oriented descriptors (G2 and G4) accounted for the two largest single-group drops (0.0187 AUC combined). Ablations had negligible internal effect (Δ ≤ 0.0020 AUC).
• Degradation reduction: Foundation-model features degraded by 0.2771 AUC under cross-cancer transfer, whereas CIT-augmented features degraded by 0.2406 AUC, reducing cross-cancer degradation by 0.0365 AUC. Mean CIT gain was larger under cross-cancer transfer than cross-site transfer (0.0446 vs. 0.0253 AUC).
• Not reported: The paper does not report pathologist-based validation of the immune clusters, external validation on non-gastrointestinal cohorts, or confidence intervals for the cross-site and reverse-direction comparisons.

Methodology in Plain English

1. Tile the slides. Each whole-slide image is split into 224 × 224 pixel tiles at 20× magnification (0.5 μm/pixel), with tissue regions selected via Otsu thresholding. Tile coordinates are kept.
2. Get frozen features. Each tile is passed through a frozen pathology foundation model — UNI2-h as the primary model, producing 1536-dimensional embeddings, and CONCH for internal comparison, producing 512-dimensional embeddings. Embeddings are L2-normalized and coordinates are normalized by slide width and height so descriptors transfer across images with different resolutions.
3. Guess which tiles are immune-rich, without labels. K-means with K = 50 (chosen by elbow curve on the training cohort) is run on up to 500 randomly sampled tiles per slide from the training cohort. Because lymphocyte-rich regions look more uniform, they fall into tighter clusters. A tightness score per cluster measures how compact it is, and the top 30% of clusters by tightness (κ = 15 of 50) are treated as immune-enriched. Each tile then gets a binary immune indicator and a continuous enrichment score.
4. Compute four families of spatial descriptors. Using only tile coordinates and the immune indicators: (a) TLS — compact immune clusters found with DBSCAN (ε = 0.02, minimum 5 points), giving membership, normalized distance to the nearest cluster centroid, and relative cluster size; (b) peritumoral margin — local spatial density from the average distance to the 20 nearest neighbors, converted to a normalized margin distance and combined with the immune score; (c) multi-scale immune density — the fraction of immune tiles among the k nearest spatial neighbors for k ∈ {10, 30, 100}; (d) immune–tumor mixing — binary entropy of the immune fraction in the 20-neighbor neighborhood plus an immune-to-tumor ratio over the 30-neighbor neighborhood.
5. Attach and aggregate. The ten descriptors are concatenated onto the frozen embedding, and the augmented tile representation is fed to an MIL aggregator (ABMIL, CLAM-SB, or TransMIL) for a slide-level MSI-H prediction.
6. Train and evaluate. Models are trained for 40 epochs with Adam (learning rate 10⁻⁴, weight decay 10⁻⁵, hidden dimension 256); the best validation-AUC checkpoint per fold is kept. TCGA-COAD (N = 301, 17.9% MSI-H) is used for training and internal five-fold stratified cross-validation. CPTAC-COAD (N = 105, 22.9% MSI-H, different institution and scanner) tests cross-site robustness, and TCGA-STAD (N = 308, 17.5% MSI-H) tests zero-shot cross-cancer transfer. External predictions are averaged across the five folds, the clustering model is fixed from the training cohort, and significance uses the paired DeLong test with Bonferroni adjustment to α = 0.0167 for three comparisons (98.33% confidence intervals).

Why This Matters

The paper addresses a well-known gap: computational pathology models for MSI-H perform strongly within a cancer type but degrade when applied across cancer types and institutions. If immune spatial organization really is organ-invariant, it offers a route to biomarker prediction that does not require collecting labeled data for every new cancer type or scanner, and that can be bolted onto existing MIL pipelines as a ten-dimensional addition rather than a new architecture.

Real-world applications

• Cross-institution deployment of MSI-H screening models — the CPTAC-COAD cross-site results (CLAM-SB 0.7989 to 0.8344, TransMIL 0.7870 to 0.8277) point toward models that hold up when a hospital uses a different scanner or staining protocol from the training site.
• Extending biomarker prediction to cancers with little labeled data — zero-shot transfer from colorectal to gastric cohorts suggests a way to offer MSI-H inference for tumor types where no local training cohort exists.
• Pre-screening to prioritize molecular testing — a computational MSI-H signal derived from routine H&E slides could help flag cases for confirmatory molecular assays.
• Reuse of existing foundation-model investments — CIT operates on frozen embeddings, so institutions already running UNI2-h or similar models can add it without retraining the feature extractor.

Industry relevance: Pathology AI vendors and digital pathology platforms must show generalization across scanners, sites, and specimen types for regulatory and commercial credibility. A method that improves transfer with only ten extra dimensions per tile, no annotations, and no target-domain data is cheap to integrate, and the authors' public release of code, pre-computed descriptors, and evaluation scripts lowers the barrier to reproducing and building on the result.

Future Directions

• Histological validation of the immune clustering. The authors state the unsupervised immune clustering lacks explicit histological validation and call for verification with pathologist annotations or cell-level segmentation.
• Non-gastrointestinal cancer types. The evaluation is restricted to COAD and STAD; the authors propose testing on TCGA-UCEC and mention endometrial and ovarian cancers to establish broader pan-cancer generalizability.
• Understanding the aggregator dependence. CIT helped ABMIL and TransMIL significantly on TCGA-STAD but not CLAM-SB after correction, and left ABMIL essentially unchanged cross-site; what determines whether a given aggregator benefits is left open.
• Mechanistic interpretation of the descriptor families. Ablation shows immune–tumor mixing and peritumoral margin matter most, but whether these should be refined, reweighted, or extended with additional immune descriptors is an open design question.

Target Audience

• Computational pathology and medical imaging researchers working on foundation models, MIL aggregation, and domain generalization.
• Computational biologists and immuno-oncology researchers interested in spatially resolved immune correlates of MSI-H status.
• Machine learning engineers building clinical biomarker prediction or digital pathology pipelines who need representations that transfer across sites and cancer types.
• Clinical informatics and regulatory teams assessing how much cross-site and cross-cancer performance loss is acceptable before a model can be deployed.
• Readers need comfort with AUC-based benchmarking and MIL terminology; the descriptor construction itself is described in sufficient detail to follow with intermediate machine learning background.

Authors’ abstract

Pathology foundation models integrated with multiple instance learning achieve competitive accuracy within single-cancer cohorts, yet cross-cancer generalization remains unresolved due to organ-specific histological and architectural differences. In this paper, we propose Conserved Immune Topology (CIT), a lightweight spatial representation for cross-cancer MSI-H prediction that augments foundation-model embeddings with biologically motivated immune descriptors. CIT uses unsupervised clustering to identify immune-associated tiles, then encodes tertiary lymphoid structures, peritumoral immune reactions, multi-scale tumor-infiltrating lymphocyte density, and immune-tumor mixing from frozen foundation-model embeddings and tile coordinates without requiring annotations or target-domain data. The proposed method was evaluated under cross-site and cross-cancer settings using CPTAC-COAD and TCGA-STAD cohorts, which introduce scanner variability, distribution shifts, and organ-specific architectural variations. Zero-shot cross-cancer transfer with CIT increased TransMIL AUC from 0.6627 to 0.7161, an absolute gain of 0.0534 (p=0.003), with consistent improvements across all three MIL aggregators. These results suggest that spatial immune topology provides potentially an organ-invariant representation for MSI-H prediction, supporting cross-cancer generalization of pathology foundation models.