PaperlayerView exact revision on arXiv

Source-linked AI summary

Input- and Feature-Space Correction for Cross-Hospital IHC Grading

Jongwoon Lee, Jhair Gallardo, Christopher Kanan

arXiv:2610.22756v1cs.CV

TL;DR

Cross-hospital differences can impair IHC grading, yet correction methods for this setting have not been systematically compared. The study benchmarks label-free corrections before and after frozen foundation-model encoding; FIX and feature-wise standardization perform strongly at their respective stages, and their combination has the highest overall mean performance.

  • Problem

    The relative performance of input- and feature-space corrections for cross-institutional IHC grading with frozen pathology foundation models remains unclear.

    Source: §1 · p. 1

  • Method

    The study compares label-free input- and feature-space corrections, including combinations, using UNI and Virchow2 across three IHC markers and six directed transfers.

    Source: §1 · p. 2

  • Results

    FIX achieved the highest mean macro-F1 among standalone input-space methods, FS led feature-space methods, and FIX+FS achieved the highest overall mean macro-F1 and QWK and lowest ordinal error.

    Source: §7 · p. 8

  • Takeaways & Limitations

    Input- and feature-space corrections provide complementary benefits for cross-institutional IHC grading without updating the foundation model or using target evaluation labels.

    Source: §6 · p. 8, §7 · p. 8

    The study covers three IHC markers, two frozen encoders, and a limited number of institution pairs, and its 12 settings reuse cohorts and images rather than providing independent external validations.

    Source: §6 · p. 8

Abstract

from arXiv · show

Pathology foundation models (FMs) are trained predominantly on hematoxylin and eosin (H&E) images but are increasingly used as frozen feature extractors for immunohistochemistry (IHC), where cross-hospital differences can impair grading. However, systematic comparisons of correction methods for cross-hospital IHC grading remain limited. To address this gap, we benchmark label-free transformations applied to input images before frozen-FM encoding and to features after frozen-FM encoding using UNI and Virchow2, three IHC markers, and six directed cross-hospital transfers, yielding 12 settings. Among the evaluated input-space methods, our proposed method FIX achieved the largest macro-F1 improvement over no correction, outperforming well-known stain normalization techniques. Among feature-space methods, a simple feature-wise standardization achieved the best improvement over no correction, outperforming CORAL. Under patient-averaged tile-level evaluation, combining FIX with feature-wise standardization improved macro-F1 and quadratic weighted kappa (QWK) by 0.181 and 0.208, respectively, and reduced ordinal error by 0.209 relative to no correction. These results show that input- and feature-space transformations can provide complementary benefits for cross-hospital IHC grading with frozen pathology FMs.

1 Introduction

Cross-institutional IHC grading can be affected by differences in tissue processing, staining, and scanners, while correction methods have not been systematically compared for frozen pathology foundation models. The study benchmarks label-free input- and feature-space corrections and reports complementary benefits from combining them.

1 Introduction

  • HER2 IHC grading contributes directly to treatment selection, while site-specific differences in processing, staining, and scanners can reduce a model’s performance at another institution.IHC supports biomarker assessment in precision oncology [Wolff et al., 2023, Modi et al., 2022].Source: §1 · p. 1
  • Pathology foundation models are predominantly trained on H&E but increasingly used as frozen feature extractors for IHC, and cross-institutional IHC correction methods remain insufficiently compared.Source: §1 · p. 1
  • The benchmark compares label-free input- and feature-space corrections using UNI and Virchow2 across three IHC markers and six directed transfers, yielding 12 settings.Source: §1 · p. 2
  • The study provides a controlled comparison of label-free input- and feature-space transformations for cross-institutional IHC grading with frozen pathology foundation models.Source: §1 · p. 2
  • FIX, based on a source-patch PCA basis and component-wise input correction, achieved the largest mean macro-F1 improvement among standalone input-space methods.Source: §1 · p. 2
  • FIX + FS achieved the highest mean performance among the evaluated input- and feature-space combinations.The comparison targets cross-institutional IHC grading with frozen pathology foundation models.Source: §1 · p. 2

2 Related Work

Prior correction methods were developed and evaluated mainly for H&E, leaving their relative performance in cross-institutional IHC grading unclear. The work studies this gap across multiple markers and frozen pathology foundation models.

2 Related Work

  • Pathology foundation models are self-supervised encoders trained on large slide collections and commonly used as frozen feature extractors; UNI is predominantly H&E-pretrained, while Virchow2’s larger corpus also includes IHC [Chen et al., 2024, Zimmermann et al., 2024].Source: §2 · p. 2
  • Input-space methods alter color or stain distributions before encoding, whereas feature-space methods align embedding statistics, with CORAL additionally matching covariance [Reinhard et al., 2001, Macenko et al., 2009, Tzeng et al., 2014, Li et al., 2018, Sun et al., 2016].The comparison keeps the pathology foundation model frozen.Source: §2 · p. 2
  • Cross-institutional adaptation of frozen pathology foundation models remains insufficiently studied for IHC, where reagents, antibody clones, staining platforms, and scanners introduce variation.Source: §2 · p. 2

3 Methods

The study tests label-free corrections before and after frozen-FM encoding, including combinations of the two stages. Input-space methods alter RGB tiles; feature-space methods alter embeddings, with corrections estimated using unlabeled target calibration data.

3 Methods

  • Corrections are compared across input and feature spaces, with the frozen encoder retained and combinations tested for complementary benefits.Labeled source tiles train the probe, unlabeled target calibration tiles estimate corrections, and held-out target tiles supply final metrics.Source: §3 · p. 2

3.1 Input-space methods

  • Reinhard and Macenko map source tiles toward target-calibration color or stain statistics, while held-out target tiles are encoded without stain normalization.Reinhard matches channel-wise LAB statistics; Macenko aligns estimated stain vectors and concentrations.Source: §3.1 · p. 3
  • FIX selectively corrects target-patch coefficients along a fixed source-patch PCA basis, using source and unlabeled target-calibration statistics.The corrected patches are reconstructed in RGB space; the source basis itself is not modified.Source: §3.1 · pp. 3–4

3.2 Feature-space methods

  • Feature-space methods align source embeddings to target calibration statistics: MS matches means, FS matches means and standard deviations, and CORAL aligns covariance.Ledoit–Wolf CORAL uses shrinkage covariance estimates for numerical stability in high-dimensional feature spaces.Source: §3.2 · pp. 4–5

4 Experimental Setup

Experiments use two institutions per marker as alternating source and target cohorts, two frozen pathology foundation models, and separate linear probes trained on labeled source data. Performance is averaged across five calibration–evaluation splits within each setting.

Data and tile annotations

  • The 7,499-tile pool covers HER2, Nectin-4, and TROP2; alternating the source–target direction and using two encoders yields 12 settings.Two pathologists independently graded tiles and resolved disagreements by consensus.Source: §4 · p. 5

Foundation-model representation

  • UNI uses a 1,024-dimensional class-token representation, while Virchow2 uses a 2,560-dimensional embedding combining class and mean patch tokens.The reported representations follow the recommended configurations for each model.Source: §4 · p. 5

Probe training

  • Separate multinomial logistic-regression probes are trained on labeled source embeddings, without target grades or held-out target data in probe training.Target calibration statistics may transform source embeddings; the trained probe is evaluated on held-out target embeddings.Source: §4 · p. 5

Metrics

  • Macro-F1, QWK, and ordinal error are reported at tile and patient-averaged tile levels, averaged over five splits per setting.Table 1 reports UNI tile-level macro-F1 means across six transfer directions.Source: §4 · p. 5, Table 1 · p. 6

5 Results

Across the evaluated methods, FIX was the strongest standalone input-space correction, FS was the strongest standalone feature-space correction, and their combination performed best overall. Patient-averaged evaluation also favored FIX+FS over no correction across the three reported metrics.

Input-space correction results

  • 0.689 mean macro-F1 made FIX the strongest standalone input-space method for UNI, and 0.609 did so for Virchow2.FIX led in five of six UNI transfer directions and four of six Virchow2 directions, but not in either Virchow2 HER2 direction; its result depended on marker, direction, and encoder.Source: §5 · p. 6, Table 2 · p. 7

Feature-space correction results

  • 0.686 mean macro-F1 for UNI and 0.677 for Virchow2 made FS the strongest standalone feature-space method for both encoders.FS outperformed CORAL; the authors note that full covariance estimation may be less stable with high-dimensional embeddings and limited target calibration data.Source: §5 · p. 6, Table 2 · p. 7

Combining input- and feature-space methods

  • 0.748 macro-F1 for UNI and 0.738 for Virchow2 were the highest overall means, both achieved by FIX+FS.For UNI, FIX+FS scored 0.748 versus 0.689 for FIX and 0.686 for FS; for Virchow2, it scored 0.738 versus 0.609 for FIX and 0.677 for FS.Source: §5 · pp. 6–7, Table 2 · p. 7

Additional analyses

  • 0.719 macro-F1, 0.744 QWK, and 0.269 ordinal error were the highest patient-averaged means, achieved by FIX+FS.Relative to no correction, FIX+FS improved macro-F1 and QWK by 0.181 and 0.208, respectively, and reduced ordinal error by 0.209.Source: Table 3 · p. 7, §5 · p. 7

6 Discussion

FIX and FS address distribution shift at different stages, and their combination retained its performance ranking under patient-averaged evaluation. The study also notes limits on the breadth and independence of its validation.

6 Discussion

  • Patient-averaged evaluation preserved the method ranking, indicating the observed benefit was not driven solely by differences in tiles contributed per patient.FIX modifies target images before frozen-FM encoding, while FS aligns feature dimensions used for probe training.Source: §6 · p. 7
  • Figure 2 compares per-setting FIX+FS and FS differences across 12 settings; positive values indicate better FIX+FS performance in all three panels.Open circles mark settings where FIX+FS performs worse than FS.Source: Figure 2 · p. 8
  • The study covers three IHC markers, two frozen encoders, and a limited number of institution pairs; its 12 settings reuse cohorts and images.The settings therefore do not constitute independent external validations; FIX gating constants were not validated on a separate marker or institution.Source: §6 · p. 8

7 Conclusion

The conclusion is that combining label-free corrections before and after frozen-FM encoding provides complementary benefits for cross-institutional IHC grading.

7 Conclusion

  • FIX+FS produced the highest overall mean macro-F1 and QWK and the lowest ordinal error across the 12 settings.Source: §7 · p. 8
  • FIX achieved the highest mean macro-F1 among evaluated standalone input-space methods, while FS led among feature-space methods.Source: §7 · p. 8
  • The paper concludes that combining input- and feature-space corrections yields complementary benefits for cross-institutional IHC grading.Source: §7 · p. 8

Ethics and Data Availability

The study used two in-house cohorts and a public cohort, and reports institutional review board approval with a waiver of informed consent.

Ethics and Data Availability

S1 Implementation details

S1 describes FIX’s drift-based component gating and the target-data handling used for calibration, evaluation, and dataset comparisons. It also records limits on patient identification and validation of the gating constants.

S1.1 FIX component-wise gating

  • FIX selects its correction pattern from the drift profile and source PCA eigenvalues, while applying stronger correction to components with concentrated source–target differences.Components with little difference receive little or no correction.Source: S1.1 FIX component-wise gating · p. 11
  • Only thresholded and weak-continuous branches were selected; HER2 activated 7–8 of 64 components, while Nectin-4 and TROP2 rescaled all 64 at mean strength 0.26.HER2’s mean strength was near 0.87; the gating constants were not validated on a separate marker or institution.Source: S1.1 FIX component-wise gating · p. 11

S1.2 Data details

  • BCI lacks patient identifiers, so original-image identifiers serve as pseudo-patient grouping units; for Institution A Nectin-4 target analysis, only 1,023 of 1,500 tiles with resolved identifiers were used.A prespecified background-pixel rule excluded 46–77% of candidate tiles, depending on marker and site.Source: S1.2 Data details · pp. 11–12
  • Target tiles are split into calibration and evaluation sets by patient, with grade distributions balanced as closely as possible and all three grades present in both sides of all 60 splits.The split is approximately 50/50.Source: S1.2 Data details · p. 11

S1.3 Source, calibration, and held-out target dataset usage

Related papers

Loading 2610.22756v1…