acceptodds
Under review as a conference paper at ICLR 2027

PixStain: Pixel-Space Diffusion with Cross-View Representation Alignment for Virtual Staining

12 views

Abstract

Virtual staining translates digital histology images between staining modalities, offering a faster and less expensive alternative to immunohistochemistry (IHC) and multiplex immunofluorescence (mIF). Existing diffusion-based virtual staining approaches generate high-quality images, but they operate in a VAE latent space that discards fine histological detail, require iterative sampling, and rely on output-space supervision alone. We observe that pathology foundation model (PFM) features provide informative spatial priors for virtual staining: their similarity maps closely align with regions of high marker expression, whereas those derived from VAE latents are dominated by noise. We therefore propose PixStain, a pixel-space diffusion model for virtual staining built on cross-view representation alignment. PixStain conditions a pixel-space Vision Transformer on multi-level PFM features of the H&E image and trains it with flow matching under x-prediction, so that a single stage of training yields the target marker image in a single inference step. Cross-view alignment matches shallow student features from spatially disjoint marker noising and H&E masking to deeper teacher features from a cleaner view. These complementary corruptions encourage the student to integrate information from both inputs. On three publicly available same-section datasets covering H&E-to-IHC, H&E-to-mIHC, and H&E-to-mIF, PixStain achieves state-of-the-art performance across pixel-level and perceptual metrics, as well as cell classification and segmentation metrics, except for pixel-level metrics on HEMIT, where it ranks second.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Staining, Pixel-Space Diffusion, Representation Learning

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_e5CjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PixStain, a pixel-space diffusion framework for virtual staining that transforms H&E images into IHC, mIHC, or mIF images. Unlike existing methods that primarily perform generation within the VAE latent space, PixStain predicts the target stained image directly in pixel space, utilizing multi-layer features extracted from a frozen pathology foundation model as conditioning. The model is trained using x-prediction-based flow matching and supports single-step inference.

Critical strengths
  1. The paper conducts experiments on three datasets, covering single-marker IHC, multiplex IHC, and multiplex immunofluorescence, and goes beyond traditional pixel-level metrics to further evaluate perceptual quality, cell classification, and cell segmentation. The method achieves strong and consistent results across these diverse evaluation metrics, thereby demonstrating significant experimental robustness and practical application value.

  2. PixStain enables single-step inference while maintaining strong performance. The paper presents ablation studies on multi-level pathology foundation model features, cross-view representation alignment, disjoint masking, and inference steps.

Critical weaknesses
  1. Most of PixStain's key components have direct counterparts in recent literature: pixel-space diffusion and x-prediction largely follow JiT; the use of a pathology foundation model to provide conditioning information for virtual staining appeared in HistDiT, and foundation-model-based virtual staining has been validated by works such as MIPHEI; meanwhile, representation alignment, shallow-to-deep feature matching, and the cleaner EMA teacher are closely related to SRA/Self-Flow. Building upon these existing modules, the paper's primary novel design appears to be the application of target-marker spatial noising and H&E masking to non-overlapping regions, thereby establishing what is termed "cross-view representation alignment." While this design is logical for the virtual staining context and empirically effective, from a methodological standpoint, it resembles a task-specific extension of Self-Flow rather than a novel principle of generative modeling or representation learning. This aspect raises concerns regarding whether the contribution meets the standards expected for ICLR.

  2. PSNR and SSIM tend to favor smooth, averaged predictions. Consequently, there is an inconsistency in the evaluation logic: if pixel-level similarity is insufficient to represent high-quality virtual staining, it should not play such a pivotal role in demonstrating the method's primary performance. Although the paper also reports perceptual and cell-level metrics, which mitigates the issue to some extent. These evaluations remain limited. For instance, the "ground truth" for cell segmentation in HER2match is not manually annotated but derived from automated segmentation using Cellpose-SAM on real IHC images; thus, its clinical or biological validity remains indirect. Admittedly, this is a major challenge in the field—one that I expect the authors will struggle to resolve during the rebuttal phase, given that truly reliable annotations are both difficult and costly to obtain.

Questions for authors

I would like to add that the aforementioned questions regarding novelty are not intended to dismiss the value of this work. I have carefully read the authors' appendix; the experimental scope is substantial and effectively demonstrates that there is still significant room for development in pathology foundation models. I will also carefully consider the comments from other reviewers when adjusting my score.

Reviewer_fmwdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes PixStain, a single step, pixel-space diffusion model for virtual staining. PixStain outperforms other methods in three publicly available virtual staining datasets in pixel-level metrics, as well as cell classification.

Critical strengths
  • The feature alignment using self-flow with disjoint matching for the student, together with multilevel PFM embedding grounding and alignment is novel in virtual staining and seems to improve performance.
Critical weaknesses
  • It is unclear why a pixel space model is needed specifically for virtual staining. Pathology image generation models have achieved very high-quality image generation while operating in the latent space [1-4], whereas for virtual staining specifically, the datasets have significant pixel-level alignment issues, as they are either produced from regions in neighboring tissue sections or even in the higher-quality datasets, like HER2Match, from the same region that is washed and restained, a process which alters the tissue, moving cells around and creating misalignments [5]. Similarly, this work evaluates virtual staining only on pixel-level metrics, which are unreliable due to data misalignment and not strongly representative of diagnostic quality, since IHC images are evaluated classified diagnostically by counting the percentage of cells or the membrane that are stained in the image [6]. For this reason, [2] and [3] evaluated generated images by their proposed clinical utility or diagnostic accuracy.

  • There are important missing references and comparisons to prior work: [2]-[4] used embeddings of PFMs to create pathology image generation foundation models, and [2] and [3] used the foundation models to perform virtual staining by using flow matching from H&E embeddings to IHC embeddings in the embedding space of a PFM.

[1]: Alfasly, Saghir, et al. "Semantic and visual crop-guided diffusion models for heterogeneous tissue synthesis in histopathology." NeurIPS 2025 [2]: Yellapragada, Srikar, et al. "Pixcell: A generative foundation model for digital histopathology images." arXiv 2025 [3]: Xiang, Jinxi, et al. "A Generative Foundation Model for Multimodal Histopathology." arXiv 2026 [4]: Duboudin, Thomas, et al. "CytoSyn: a Foundation Diffusion Model for Histopathology--Tech Report." arXiv 2026 [5]: Klöckner, Pascal, et al. "GANs vs. Diffusion Models for virtual staining with the HER2match dataset." MICCAI Workshop on Deep Generative Models 2025 [6]: Klöckner, Pascal, et al. "H&E to IHC virtual staining methods in breast cancer: an overview and benchmarking." Npj digital medicine 2025

Questions for authors
  • Why is a pixel-level alignment important specifically for virtual staining? Is it specifically important for H&E to mIF virtual staining?
  • Is it possible to provide metrics that evaluate the clinical utility of virtually stained images?
Reviewer_SBfQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PixStain is a conditional pixel-space diffusion model for translating H&E images into single-marker or multiplex staining images. It combines a Vision Transformer with multilevel features from a frozen H-Optimus-0 encoder and uses flow matching with clean-image prediction. Its principal training contribution is cross-view representation alignment: a student receives spatially disjoint target-image noising and H&E masking, while an exponential-moving-average teacher receives a cleaner target and unmasked H&E. Shallow student representations are aligned with deeper teacher representations. An additional perceptual loss is used for single-marker IHC, and inference uses one Euler step.

The method is evaluated on three same-section paired datasets: HER2match, HEMIT, and ORION-CRC. Comparisons cover regression, GAN, and diffusion approaches, using image-reconstruction, perceptual, and downstream cell-analysis metrics. PixStain reports the strongest aggregate results in most categories, with MIPHEI retaining better pixel-level performance on HEMIT. Ablations investigate conditioning features, alignment, masking, perceptual loss, and inference steps.

Critical strengths
  1. The method integrates pathology-specific features into a pixel-space generator and adapts representation alignment to paired image translation through complementary corruption of the source and target. Importantly, the manuscript does not rely entirely on comparisons against unrelated architectures: Tables 4 and 5 on page 9 compare foundation encoders, feature-layer aggregation, and alternative H&E masks. The disjoint-mask configuration improves PSNR, SSIM, PCC, Dice, and AJI relative to the unmasked configuration, while matching its LPIPS. These comparisons provide meaningful evidence for the proposed conditioning and corruption choices. The demonstrated single-step inference is also a potentially useful.

  2. Using same-section pairs is an important design choice for a task in which cellular correspondence matters. The evaluation spans single-marker IHC and multiplex targets, and includes cell-analysis outcomes rather than relying exclusively on image similarity. For example, on HEMIT, PixStain improves average cell-classification F1 from MIPHEI’s 0.716 to 0.772, despite slightly worse pixel-level reconstruction. This provides evidence that the method’s value is not captured by PSNR alone and is consequential to the intended downstream application. The per-marker appendix also shows heterogeneous performance.

Critical weaknesses
  1. The central explanation for the gains is not sufficiently isolated from stronger conditioning and perceptual supervision. The paper argues that VAE compression is a principal limitation of earlier methods, but the similarity maps in Figure 1 do not directly establish how much task-relevant information is irreversibly lost. A matched pixel-versus-latent comparison, holding conditioning and losses fixed, is missing from the reported ablations.

The attribution to cross-view alignment is also incomplete. In Table 3, adding alignment to multilevel PFM conditioning improves HER2match’s pixel-level metrics, but LPIPS changes from 0.423 to 0.444 and Dice from 0.520 to 0.518. The subsequent addition of perceptual loss produces the substantial improvements to 0.276 LPIPS and 0.576 Dice. The missing PFM-plus-perceptual-loss control without alignment is therefore important.

Finally, combining Equations 6 and 15 shows that one-step inference returns the network’s clean-image prediction at �0��0�. This is valid, but does not itself establish an advantage over direct image regression using the same features, backbone, and image losses. These controls matter because the contribution should be distinguished from the benefits of a strong pathology encoder and perceptual supervision.

  1. The validation does not establish how robust the improvements are across independent specimens or how faithfully individual markers are recovered. The large tile counts obscure a small number of independent specimens: HER2match contains 17 cases, HEMIT eight patients, and ORION-CRC uses only two test WSIs. HEMIT’s patient-disjoint split is explicit, whereas HER2match is described through tile counts without equivalent patient-level detail. The reported comparisons lack uncertainty estimates or repeated-training results, making small differences—such as ORION-CRC AUROC 0.886 versus 0.876—difficult to assess beyond the particular split.

The downstream metrics are useful but remain indirect measures of staining fidelity. HER2match segmentation compares two sets of automatically generated Cellpose-SAM masks, not expert reference annotations. Multiplex classification uses a learned classifier on predicted marker-intensity vectors, which can exploit cross-marker associations without establishing accurate reconstruction of each marker. Indeed, the appendix reports PD-L1 PCC of 0.221 and AUPRC of 0.051; prevalence and direct marker-agreement analyses are needed to interpret this result. Specimen-level paired results and marker-specific fidelity checks would substantially strengthen confidence in the claimed biological utility.

Questions for authors
  1. What are the patient and slide counts in each HER2match partition, and are all partitions patient-disjoint? For ORION-CRC, are the improvements consistent across both test slides? Please provide available specimen-level paired results and indicate whether training-seed variability was assessed.

  2. Can the authors provide the HER2match result using multilevel PFM conditioning and LPIPS without representation alignment? Does PixStain retain an advantage over direct prediction using the same conditioning, backbone, and image losses?

  3. Are the downstream logistic-regression classifiers refitted separately for each generator? Can the authors report direct predicted-versus-measured marker agreement using a fixed evaluation rule, together with marker prevalence - particularly for PD-L1? This would help distinguish phenotype-predictive representations from faithful marker reconstruction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.