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

Jongwoon Lee; Jhair Gallardo; Christopher Kanan

cs.CVOct 15, 2026

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

Authors: Jongwoon Lee, Jhair Gallardo, Christopher Kanan

Organizations: Department of Computer Science, University of Rochester, Rochester, NY · Rochester Institute of Technology, Rochester, NY

Abstract

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.

Figures & tables

Figure 1: Overview of the evaluated label-free transformations. Input-space methods act before frozen-FM encoding, whereas feature-space methods act on the resulting embeddings.

Method	Correction stage	HER2		Nectin-4		TROP2
Method	Correction stage	$\mathrm{A}\to\mathrm{BCI}$	$\mathrm{BCI}\to\mathrm{A}$	$\mathrm{A}\to\mathrm{B}$	$\mathrm{B}\to\mathrm{A}$	$\mathrm{A}\to\mathrm{B}$	$\mathrm{B}\to\mathrm{A}$	Mean
No correction	None	0.356	0.564	0.412	0.549	0.817	0.699	0.566
Reinhard	Input	0.401	0.606	0.465	0.525	0.726	0.686	0.568
Macenko	Input	0.306	0.391	0.533	0.511	0.410	0.471	0.437
PCA projection	Input	0.288	0.566	0.669	0.644	0.831	0.816	0.636
FIX	Input	0.605	0.335	0.689	0.781	0.881	0.844	0.689

Table 1: Tile-level macro-F1 for label-free correction methods using UNI. Values are means over five target calibration–evaluation splits. Mean denotes the average across the six transfer directions. Bold indicates the numerically highest value.

Method	Correction stage	HER2		Nectin-4		TROP2
Method	Correction stage	$\mathrm{A}\to\mathrm{BCI}$	$\mathrm{BCI}\to\mathrm{A}$	$\mathrm{A}\to\mathrm{B}$	$\mathrm{B}\to\mathrm{A}$	$\mathrm{A}\to\mathrm{B}$	$\mathrm{B}\to\mathrm{A}$	Mean
No correction	None	0.578	0.170	0.526	0.550	0.893	0.461	0.530
Reinhard	Input	0.633	0.375	0.385	0.480	0.752	0.489	0.519
Macenko	Input	0.501	0.409	0.270	0.505	0.698	0.367	0.459
PCA projection	Input	0.397	0.147	0.598	0.666	0.781	0.616	0.534
FIX	Input	0.625	0.153	0.605	0.750	0.847	0.676	0.609

Table 2: Tile-level macro-F1 for label-free correction methods using Virchow2. Values are means over five target calibration–evaluation splits. Mean denotes the average across the six transfer directions. Bold indicates the numerically highest value.

Method	Macro-F1 $\uparrow$	QWK $\uparrow$	Ordinal error $\downarrow$
No correction	0.539	0.536	0.478
FIX	0.652	0.673	0.332
FS	0.662	0.675	0.329
FIX + FS	0.719	0.744	0.269

Table 3: Patient-averaged tile-level performance. Each identified patient received equal total weight. For BCI, released original-image identifiers served as grouping units because patient identifiers were unavailable. Values are means across the 12 settings.

Panel of Figure 2 — Figure 2: Per-setting effect of adding FIX to FS, under patient-averaged tile-level evaluation. Each point shows the difference between FIX+FS and FS in the mean over the five calibration–evaluation splits, using the same weighting as Table 3 . Ordinal error is lower-is-better, so its panel plots FS $-$ FIX+FS; positive values therefore indicate better performance with FIX+FS in all three panels. The three panels share the same horizontal scale. The solid vertical line marks zero, and the dashed line marks the mean across the 12 settings. Open circles indicate settings in which FIX+FS performs worse than FS. V2 denotes Virchow2 and NEC4 denotes Nectin-4.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Condition	Target-calibration use	Source / probe-training representation	Held-out target representation
No correction	None	Unmodified source images encoded by the frozen FM	Unmodified target images
PCA projection	None; the basis is fitted on source patches only	PCA-reconstructed source images encoded by the frozen FM	PCA-reconstructed target images
FIX	Estimate coefficient statistics from source and unlabeled target calibration RGB patches	PCA-reconstructed source images, without coefficient editing	FIX-corrected target images
MS	Mean of the unmodified target calibration features	Unmodified source features recentered to that mean	Unmodified target images
FS	Per-dimension mean and standard deviation of the unmodified target calibration features	Unmodified source features standardized to those statistics	Unmodified target images
PCA projection+MS	Mean of the PCA-reconstructed target calibration features	PCA-reconstructed source features recentered to that mean	PCA-reconstructed target images, unshifted

Appendix

Table S1: Training and evaluation procedures for the evaluated conditions. All conditions use the same frozen pathology FM and the same source grades. Held-out target evaluation data are used neither for correction estimation nor for probe training.

Domain Adaptation

Domain adaptation transfers machine-learning models across changing data distributions, often using unlabeled or limited target-domain examples.

See all 253 papers

UNIStainNet: Foundation-Model-Guided Virtual Staining of H&E to IHC
Mar 13, 2026Jillur Rahman Saurav, Thuong Le Hoai Pham, Pritam Mukherjee +3Virtual Staining Computational Pathology
Department of Computer Science, University of Texas at Arlington, Arlington, Texas, USA · Department of Radiology, St. Jude Children’s Research Hospital, Memphis, Tennessee, USA · Department of Pathology, St. Jude Children’s Research Hospital, Memphis, Tennessee, USA
Histopathology Image Normalization via Latent Manifold Compaction
Feb 27, 2026Xiaolong Zhang, Jianwei Zhang, Selim Sevim +3Computational Pathology Domain Generalization
Knight Cancer Institute, Oregon Health and Science University, Oregon, USA · Brenden-Colson Center for Pancreatic Care, Oregon Health and Science University, Oregon, USA
Can You Trust Frozen Hematology Foundation Models under Acquisition Shift?
Aug 25, 2026Jai Kumar Sharma, Peeyush TapadiyaWhite Blood Cell Classification Medical Imaging Foundation Models
Virginia Tech · Accenture

Mar 13, 2026J/K move · Enter open · S save

UNIStainNet: Foundation-Model-Guided Virtual Staining of H&E to IHC

Jillur Rahman Saurav, Thuong Le Hoai Pham, Pritam Mukherjee, Paul Yi+2

Department of Computer Science, University of Texas at Arlington, Arlington, Texas, USA · Department of Radiology, St. Jude Children’s Research Hospital, Memphis, Tennessee, USA · Department of Pathology, St. Jude Children’s Research Hospital, Memphis, Tennessee, USA

Virtual Staining Computational Pathology Image-to-Image Translation

Open paper AskPDF arXiv

Virtual immunohistochemistry (IHC) staining from hematoxylin and eosin (H&E) images can accelerate diagnostics by providing preliminary molecular insight directly from routine sections, reducing the need for repeat sectioning when tissue is limited. Existing methods improve realism through contrastive objectives, prototype matching, or domain alignment, yet the generator itself receives no direct guidance from pathology foundation models. We present UNIStainNet, a SPADE-UNet conditioned on dense spatial tokens from a frozen pathology foundation model (UNI), providing tissue-level semantic guidance for stain translation. A misalignment-aware loss suite preserves stain quantification accuracy, and learned stain embeddings enable a single model to serve multiple IHC markers simultaneously. On MIST, UNIStainNet achieves state-of-the-art distributional metrics on all four stains (HER2, Ki67, ER, PR) from a single unified model, where prior methods typically train separate per-stain models. On BCI, it also achieves the best distributional metrics. A tissue-type stratified failure analysis reveals that remaining errors are systematic, concentrating in non-tumor tissue. Code is available at https://github.com/facevoid/UNIStainNet.

Abstract

Domain Adaptation

Explore similar work

UNIStainNet: Foundation-Model-Guided Virtual Staining of H&E to IHC

Histopathology Image Normalization via Latent Manifold Compaction

Can You Trust Frozen Hematology Foundation Models under Acquisition Shift?