acceptodds
Under review as a conference paper at ICLR 2027

Beyond Spatial Correspondence: Optical-Density-Guided Marginal and Relational Learning for Weakly-Paired H&E-to-IHC Virtual Staining

10 views

Abstract

Weakly paired H&E-to-IHC virtual staining learns from adjacent tissue sections, where deformation, tissue loss, and expression heterogeneity weaken local correspondence. Spatial co-location and visual similarity do not necessarily imply compatible staining states, making them insufficient on their own for organizing stain-consistent supervision. Optical density (OD), widely used to quantify immunohistochemical (IHC) staining intensity, provides complementary evidence of stain expression. We propose PGLStain, which uses continuous OD compatibility not merely as an intensity-matching target, but as a signal for conditioning feature relations and allocating local supervision. Protein-Expression-Aware Correspondence Calibration (PECC) combines feature-context and OD-calibrated graphs, assigns relative weights to spatially proposed candidate pairs, and aligns permutation-invariant OD-state graph summaries. To complement candidate-based supervision with image-level expression constraints, Marginal–Relational Stain Alignment (MRSA) aligns unordered OD marginals and uses unbalanced Gromov–Wasserstein transport to compare intra-image expression-state relations under relaxed node-mass constraints, without fixed cross-section node assignments. Together, these objectives integrate stain-calibrated local supervision with correspondence-free expression-state learning, without additional inference-time computation. Experiments on public H&E-to-IHC benchmarks demonstrate competitive perceptual quality and quantitative staining agreement. Code are avaliable at https://anonymous.4open.science/r/PGLStain-anonymous-AC00anonymous repository

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Weakly Paired Image Translation, Virtual Staining, Computational Pathology, Graph Contrastive Learning, Optimal Transport, Unbalanced Gromov–Wasserstein, Optical Density

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FWA3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses weakly-paired H&E→IHC virtual staining, where adjacent-section references suffer from deformation, tissue loss, and expression heterogeneity. The key idea is to use continuous (focal) optical density (OD) not only as an intensity target but as a signal for (i) calibrating local contrastive supervision (PECC: dual feature/OD-calibrated graph context encoding, OD-weighted candidate pairs, and permutation-invariant OD-graph summary alignment) and (ii) constraining regional expression composition and intra-image expression-state relations without cross-section node assignments (MRSA: quantile-based marginal alignment of block-mean FOD + entropic unbalanced Gromov–Wasserstein relational matching with a detached coupling). All auxiliary objectives are training-only. Experiments on MIST (ER/PR/HER2/Ki67) and ACROBAT-ER use 8 metrics plus a Fiji-based case-level quantification (H-score, positive nuclei/area ratios with ICC(A,1)); ablations cover the four objectives, graph topology vs. candidate weighting, alternatives to the relational term (Direct UOT / MMD / BGW), and sensitivity to ρ, ε, and loss weights.

Critical strengths
  1. The two challenges (ambiguous local correspondence; regional heterogeneity under imperfect correspondence) are stated precisely and map cleanly onto PECC and MRSA. The explicit stance that OD agreement is not correspondence ground truth is intellectually honest and correctly framed throughout.
  2. Using one OD kernel for graph calibration and cross-image candidate weighting, plus quantile-marginal and UGW-relational alignment built from the same FOD statistics, gives the method a unified design rather than a bag of losses. The UGW formulation with prescribed tissue-coverage masses, squared-mass-normalized detached relational loss, and gradient flow restricted to the virtual-side relation matrix is carefully worked out.
  3. The eight-metric suite is unusually complete for this task: paired perceptual fidelity (PHV Avg, DISTS), feature-distribution similarity in both generic and stain-specialized foundation-model spaces (FID, PathFID–StainNet), and three complementary staining-agreement measures (HistCorr, DAB–KL, per-image mIoD PearsonR plus dataset-level bias). This is complemented by the Fiji-based case-level quantification (H-score, positive nuclei/area ratios, ICC(A,1) over 55 pooled case groups), which is the evaluation closest to actual pathology practice and is rarely reported in this literature. Together these cover pixel-, distribution-, stain-, and case-level evidence.
Critical weaknesses
  1. PGLStain does not lead on 6 of 8 metrics on ACROBAT-ER: HistCorr 0.6222 vs SIMGAN 0.6466; FID 8.03 vs PSPStain 7.65; and most strikingly |Bias mIoD| 6.326 vs USIGAN 0.344 (~18× worse) and vs SIMGAN 2.301. The main text devotes two sentences to this ("competitive performance with dataset- and metric-dependent advantages") while the abstract claims "competitive ... quantitative staining agreement" broadly. The paper should (a) diagnose why — e.g., ACROBAT's larger WSI-level heterogeneity, the 1024→512 resize altering FOD statistics, or the fixed ρ=2/σ_OD=0.1 being mis-tuned for this distribution — and (b) temper the abstract. As written, a reader of the abstract and Table 1 (top half) would get a misleading impression.
  2. Key hyperparameters of the OD kernel lack sensitivity analysis: σ_OD=0.1 appears in the graph calibration (Eq. 1), candidate weighting (Eq. 14), and relation matrices (Eq. 6), yet only ρ, ε, and the λ's are swept. Given that σ_OD controls what "compatible staining" means relative to the empirical FOD range, a sweep of σ_OD and of the feature threshold τ=0.6 seems at least as important as the ε sweep. Relatedly, the hard-gating threshold δ_hard=0.12024 was calibrated on 128 validation pairs (A.1.5) — a mild form of validation-set leakage into a reported ablation; training data should be used instead.
  3. The Fiji-based evaluation is the most clinically meaningful evidence in the paper, but it is only reported on MIST-ER. Extending it to at least one additional stain would strengthen the central claim — HER2 is the most decision-relevant (semi-quantitative 0/1+/2+/3+ scoring), and Ki67 has low positive fractions where marginal alignment may behave differently. This is a suggestion for strengthening rather than a flaw in what is reported.
  4. On re-reading, the genuinely new elements are (a) continuous OD compatibility as a shared calibration signal across graph edge gating, context encoding, and cross-image candidate weighting — ASP-style adaptive weighting uses feature similarity only — and (b) GW-relational matching of expression-state descriptors with relaxed marginals, which is distinct from USIGAN's unbalanced feature-plan transport and from PSPStain's image-level FOD supervision. However, the paper does not articulate this boundary explicitly: the related-work section lists the predecessors but never states what exactly PECC/MRSA add beyond "graph context + OD" or "unbalanced transport." A short paragraph making the above distinctions precise would preempt the "incremental combination" reading.
Questions for authors
  1. Why is |Bias mIoD| on ACROBAT-ER ~18× worse than USIGAN (6.326 vs 0.344) while PearsonR is the best? Is this a systematic global intensity offset in the generated images, and if so, why don't the marginal/FOD objectives correct it?
  2. Please report sensitivity to σ_OD and τ. How were σ_OD=0.1, τ=0.6, σ_p+=0.25 chosen — on training or validation data?
  3. For the Fiji evaluation: the StarDist "Versatile (H&E nuclei)" model is applied to IHC and virtual-IHC images. How sensitive are the case-level ICCs to the detector? Does a detector mismatch penalty affect all methods symmetrically?
  4. The fixed UGW budget (6 linearizations × 30 Sinkhorn iterations) yields an approximate coupling. How large is the approximation error relative to a converged solution, and does a larger budget change rankings?
  5. ρ=2 is used for ACROBAT and all MIST stains. Was it selected on MIST-ER only? Table 8's non-monotonicity suggests per-dataset tuning risk.
Reviewer_Qmc2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PGLStain, an optical-density-guided framework for weakly paired H&E-to-IHC virtual staining. To address unreliable spatial correspondence between adjacent tissue sections, it introduces two complementary components. PECC uses OD-conditioned graph representations and continuous OD-based weighting to improve local supervision. MRSA combines regional expression distribution matching with unbalanced Gromov-Wasserstein transport to capture staining relationships without requiring explicit spatial correspondence. Experiments on two breast dataset MIST and ACROBAT demonstrate its competitive performance.

Critical strengths

(1) The paper addresses an important challenge in weakly paired virtual staining, where spatial proximity does not necessarily imply reliable staining correspondence. (2) PECC integrates optical-density compatibility into both graph-based feature representations and candidate supervision weighting. This provides a meaningful alternative to relying solely on spatial proximity or feature similarity. (3) The ablation studies include the contributions of individual objectives, graph construction, candidate weighting, etc. I think it is relatively all-round.

Critical weaknesses

(1) Although PGLStain performs well on MIST, its advantages on ACROBAT are less consistent. In Table 1, several competing methods outperform PGLStain on metrics such as FID, DISTS, and PathFID. This limits the strength of the generalization claims. (2) OD compatibility is calculated using generated and weakly paired reference IHC images. When staining predictions are inaccurate or adjacent sections exhibit substantial expression differences, OD similarity may not reliably indicate correspondence quality. (3) Although the paper includes many automated quantification. Additional expert assessment or independent biological validation would better establish whether the generated staining patterns reliably reflect actual biomarker expression. (4) Experiments are conducted exclusively on breast cancer datasets, without validation on other cancer types. Although the results demonstrate promising performance, the generalizability of the proposed method across different tissue morphologies and biomarker expression patterns remains unclear.

Questions for authors

See weaknesses. Further, (1) Could the authors provide results from multiple random seeds, particularly on MIST-ER and ACROBAT-ER, to demonstrate the stability and statistical significance of the reported improvements? (2) How does the OD-based correspondence weighting behave when the generated staining is initially inaccurate or when adjacent tissue sections exhibit substantial expression heterogeneity?

7 comments
AuthorsOpenReview ↗

Detail Respond to Weakness (1): Updated evaluation on limited performance on ACROBAT and Data Quality

On the curated ACROBAT-lite ER evaluation, PGLStain (ρ=1\rho=1) achieves the best values on six of eight metrics and the second-best PathFID. Below, we explain the data-quality findings, report the updated comparison, and show both tested marginal penalties.

1. What did the data audit reveal?

We identified black registration boundaries and white or near-white, tissue-deficient patches in the original WSI-derived data. Near-zero intensities produce large values under the logarithmic OD transform, potentially introducing non-tissue signals into DAB-derived supervision; tissue-deficient regions also provide limited expression information.

This sensitivity is visible in a fixed-prediction development-set check: excluding 22 reference images with more than 1% pure-black pixels changed PearsonR from 0.0823 to 0.7589 for ρ=2\rho=2 (1,169 to 1,147 pairs). The predictions were unchanged; this check shows how strongly these references can affect the reported metric. Details are in Table 4 of our response to FWA3, “Detail Respond to Q4 and Q5.”

2. Updated comparison on curated ACROBAT-lite ER

The curated dataset contains 4,422 training and 1,000 held-out pairs from 24 and 6 disjoint case-ID groups, respectively. PGLStain uses ρ=1\rho=1, ε=0.05\varepsilon=0.05, and a fixed epoch-100 checkpoint evaluated at 512×512512\times512.

Table 1. Updated ACROBAT-lite ER results. Best and second-best values are bold and underlined‾\underline{\text{underlined}}. PHV denotes PHV-Avg; abs. bias is absolute mIoD bias. PathFID uses StainNet features.

MethodHistCorr ↑PHV ↓DISTS ↓FID ↓PearsonR ↑Abs. bias ×1,000 ↓DAB–KL ↓PathFID ↓
CUT0.52970.57700.256352.620.659626.3880.6337463.03
CycleGAN0.52920.57720.255266.640.637530.1550.3781531.76
UNSB0.47720.63530.284470.010.591223.8110.9714631.47
PyramidP2P0.36910.74460.4981184.220.544541.0280.54111759.60
ASP0.42860.69130.3209173.390.857125.8680.61921600.36
PPT0.46580.57480.269052.100.810929.4120.6442486.65
PSPStain0.11420.58710.249054.610.8603‾\underline{0.8603}20.4670.5075873.03
MDCL0.6069‾\underline{0.6069}0.56590.253946.780.840624.7410.2555‾\underline{0.2555}427.82
SIMGAN0.40240.61030.280763.810.691728.9790.6209655.89
USIGAN0.50390.5597‾\underline{0.5597}0.2452‾\underline{0.2452}47.390.801323.643‾\underline{23.643}0.3835437.84
M2PLGAN0.54630.57430.250443.76‾\underline{43.76}0.788424.2520.3690449.24
PGLStain, ρ=1\rho=10.61470.55860.242342.290.876124.2100.2552431.41‾\underline{431.41}

PGLStain leads HistCorr, PHV, DISTS, FID, PearsonR and DAB–KL. MDCL retains the lowest PathFID, and PSPStain the lowest absolute bias.

3. Does the MIST–ER default transfer unchanged?

The default ρ=2\rho=2 was selected through the MIST–ER validation sweep. We report both completed ACROBAT-lite configurations on the same 1,000 held-out pairs, with ε=0.05\varepsilon=0.05, epoch-100 checkpoints, and identical evaluation preprocessing.

Table 2. Marginal-penalty comparison. Better values are in bold.

SettingHistCorr ↑PHV ↓DISTS ↓FID ↓PearsonR ↑Abs. bias ×1,000 ↓DAB–KL ↓PathFID ↓
ρ=1\rho=10.61470.55860.242342.290.876124.2100.2552431.41
ρ=2\rho=20.56890.55550.247645.570.868524.4830.3031434.39

Here, ρ=1\rho=1 performs better on seven metrics, while ρ=2\rho=2 gives lower PHV. Thus, the six-metric lead in Table 1 applies specifically to ρ=1\rho=1, not automatically to the transferred default.

In summary, the updated ACROBAT-lite results support PGLStain's applicability beyond MIST, with advantages spanning perceptual and staining-agreement metrics. The data audit identifies a concrete sensitivity of OD-based evaluation, while the two-configuration comparison shows that the preferred marginal penalty need not transfer unchanged across datasets.

AuthorsOpenReview ↗

OD Weighting Under Prediction Errors and Section Heterogeneity

Thank you for highlighting this ambiguity. OD disagreement can arise from either an inaccurate prediction or a genuine expression difference between adjacent sections. These situations have different implications for supervision, and OD alone cannot distinguish them.

1. When does downweighting help, and when can it be misleading?

  • Genuine section-to-section expression differences: visually similar regions need not have identical staining. Downweighting an expression-incompatible candidate reduces the influence of reference expression that may be inappropriate for that generated region. Here, low OD agreement does not necessarily mean incorrect anatomical correspondence; it means that strong local expression supervision may be inappropriate.
  • Inaccurate generated staining: a useful reference candidate can also receive a lower weight because the prediction is wrong. This can weaken useful supervision. Our weighting mechanism does not explicitly identify or resolve this case.

The weight modulates only the candidate contrastive objective; it does not discard the region or disable its other losses. The base objectives remain active, while MRSA uses regional expression distributions and relations without fixed cross-section assignments.

2. Does OD weighting provide useful supervision despite this ambiguity?

The reported candidate-weighting ablation addresses its end-to-end utility. With the dual graph representation and remaining objectives fixed, continuous OD weighting achieves better values than both uniform and feature-similarity weighting on all six reported metrics.

Table 1. Candidate-weighting ablation on MIST–ER. PathFID uses StainNet features. Best values are in bold.

WeightingHistCorr ↑DISTS ↓FID ↓PearsonR ↑DAB–KL ↓PathFID ↓
Uniform0.72210.208530.490.93930.2813165.53
Feature similarity0.73200.213130.750.93930.3064166.80
Continuous OD0.74080.199429.350.94550.2391145.15

In particular, DISTS improves from 0.2131 to 0.1994, and PathFID from 166.80 to 145.15, relative to feature weighting. These results support staining compatibility as useful information beyond visual similarity when determining supervision strength.

3. How sensitive is the supervision to changes in prediction or reference staining?

We additionally tested 64 fixed MIST–ER training pairs at epoch-20 and epoch-100 Full checkpoints. We independently increased or decreased local DAB concentration by 25% in either the generated image or the reference, keeping the other image unchanged. At each frozen checkpoint, OD and feature weighting were compared using the same perturbed inputs, graph context, and remaining objectives, without updating model parameters.

We measured the relative change in the total loss gradient with respect to the generated image. Across the eight checkpoint–perturbation conditions, the largest difference between the two weighting rules' mean full-image response magnitudes was 0.084 percentage points. Within the edited region, however, the mean candidate-loss gradient-response norm was 1.8–6.8% higher with OD weighting. Thus, the overall mean responses were comparable, but local responses were not identical.

The supported conclusion is that OD compatibility complements feature similarity when weighting weakly paired supervision. Its benefit is demonstrated by improved final outcomes under the evaluated protocol. The perturbation analysis characterizes immediate sensitivity to staining discrepancies; it does not establish that OD identifies their cause or guarantees recovery from incorrect predictions.

AuthorsOpenReview ↗

Response to Q1: Does the improvement persist across training seeds?

On MIST–ER, PGLStain achieves better three-seed mean scores than M2PLGAN on all eight metrics. In each additional training seed (1108 and 2027), it improves seven of eight metrics. These are independently trained models, not bootstrap resamples of fixed predictions.

We completed two additional runs per method using training seeds 1108 and 2027. Both methods use the same 4,153 training pairs and 1,000 validation pairs, with fixed epoch-100 checkpoints evaluated at 512×512512\times512. PGLStain retains ρ=2\rho=2 and ε=0.05\varepsilon=0.05. We verified identical validation inputs and references, metric implementations, and StainNet weights for these new comparisons. Together with the seed-42 references, Table 1 reports the mean and sample standard deviation across three training runs.

Table 1. Comparison across training seeds 42, 1108, and 2027. Bold indicates the better mean, not statistical significance. PathFID uses StainNet features.

MetricM2PLGANPGLStain
HistCorr ↑0.7106 ± 0.01380.7229 ± 0.0121
PHV Avg ↓0.5455 ± 0.01250.5315 ± 0.0081
DISTS ↓0.2143 ± 0.00300.2091 ± 0.0018
FID ↓30.46 ± 0.3830.44 ± 0.19
PearsonR ↑0.9200 ± 0.00180.9406 ± 0.0034
Absolute mIoD bias ×1,000 ↓5.307 ± 1.2083.357 ± 1.055
DAB–KL ↓0.5205 ± 0.02740.3056 ± 0.0299
PathFID ↓160.74 ± 0.43160.20 ± 2.08

The clearest repeated improvements concern staining agreement, rather than FID alone. Mean DAB–KL decreases by 41.3%, absolute mIoD bias decreases by 36.7%, and PearsonR increases from 0.9200 to 0.9406. PGLStain improves PHV, DISTS, PearsonR, absolute bias, DAB–KL, and PathFID in both additional seeds. M2PLGAN has better HistCorr at seed 1108 and better FID at seed 2027.

Reviewer_WHM3OpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Weakly paired H&E and IHC stained images always face incompatible staining stats. However, previous methods use the visual similarity to align these two kind images leading to mistakes generated image. To solve the problem, the paper uses optical density as informative signal for local supervision.At the same time, this paper proposes stain-calibrated correspondence learning to address unreliable spatial alignment and correspondence-free distribution learning to preserve regional expression distributions and relationships without requiring explicit spatial correspondence.

Critical strengths
  1. The motivation of this paper is clear and reasonable. The authors identify the limitations of spatial correspondence in weakly paired H&E-to-IHC virtual staining and provide a meaningful motivation for using optical density to improve local supervision and regional expression alignment.

  2. The experimental evaluation is relatively comprehensive. The authors compare their method with several existing approaches on different datasets and provide detailed ablation studies to evaluate the contributions of different components. The additional pathology-related quantitative evaluations also help demonstrate the effectiveness of the proposed method.

Critical weaknesses
  1. The motivation behind each module is not clearly explained, which makes the overall methodology difficult to follow. The authors introduce several graph structures, alignment strategies, and loss functions, but I find it hard to understand why all these components are necessary and how they work together. Although the overall motivation is reasonable, the connection between the research problem and each module design is not sufficiently clear.

  2. The proposed method seems to be a combination of many existing techniques rather than a well-motivated framework. Although the authors provide ablation studies, the individual contribution of some components is still not very convincing, especially when certain modules bring only limited improvements. I would like to see a clearer explanation of why these components are designed in this way and whether a simpler framework could achieve similar results.

Questions for authors

Please clearly explain the motivation behind the design of each module and reorganize Figure 2 to make the overall workflow easier to understand.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.