acceptodds
Withdrawn from ICLR 2027

CellDisc: Boosting Virtual Staining via Cell Segmentation-Guided Discrimination

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Virtual staining aims to generate histochemically stained images from unstained or differently stained tissue images, yet existing methods primarily rely on image-level objectives without explicitly modeling cellular structures that are fundamental to histopathological interpretation. Moreover, rigorous evaluation remains challenging because commonly used datasets often rely on consecutive tissue sections, resulting in imperfect cell-level correspondence between source and target images. To address these limitations, we propose CellDisc, a simple yet effective cell segmentation-guided discrimination framework that introduces dense cellular supervision into adversarial learning. Instead of conventional binary discrimination, CellDisc jointly models cell localization and directional flow fields, encouraging the preservation of cellular structure and spatial organization. As a plug-and-play discriminator enhancement, CellDisc can be integrated into diverse paired GAN baselines without modifying their generators. We further establish three strongly paired same-section datasets covering H&E-to-P63, H&E-to-P63&CK, and FFPE-to-H&E transformations, enabling unified reference-based evaluation of pixel-wise, perceptual, cell-level, and stain-specific fidelity. Extensive experiments across multiple baselines and staining tasks demonstrate the broad compatibility and effectiveness of CellDisc. When integrated with P2P-res9, CellDisc achieves competitive or superior performance compared with recent GAN- and diffusion-based approaches, particularly in perceptual and cell-level fidelity. Ablation studies further demonstrate the benefit of directional flow fields and their complementarity with perceptual supervision, while evaluations using independent cell segmentation models consistently support the improvements in cellular fidelity. Finally, downstream cancer classification demonstrates that CellDisc preserves diagnostically relevant information across different classifier architectures. Overall, our results establish explicit cellular modeling as an effective and compatible strategy for improving structural fidelity in virtual staining. Code and data will be made publicly available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Staining, GANs, Cell Segmentation Discrimination.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YfcbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces CellDisc, a virtual staining framework for computational pathology that addresses pathological semantic degradation, a limitation of existing methods that produce visually convincing but clinically inaccurate images. This is addressed by introducing dense cellular supervision into adversarial learning.

Critical strengths

None

Critical weaknesses

I don't see much interest in the paper outside the computational pathology field. The proposed solution adds application-specific loss terms to the learning problem, which don't introduce novelty to the broader algorithms field.

The use of the DAB channel, the counting of stained cells, or downstream tasks have already been considered [A]. These works should be better put into context.

The BCI dataset is known to contain many artifacts, such as doubling of the image sections or dark/black shade. [B] manually assessed all tiles and cleaned the dataset, resulting in the BCI-clean version. Why only the original dataset, and not the cleaned version? Why not other datasets, such as the HER2match dataset [B]. So, the assumption that there are no publicly available datasets is questionable.

Citations are almost all in the wrong format (citet vs citep).

References [A] Peng, Q. et al. Advancing H&E-to-IHC virtual staining with task-specific domain knowledge for HER2 scoring. In Medical Image Computing and Computer Assisted Intervention—MICCAI, 2024.

[B] Klöckner, P., Teixeira, J., Montezuma, D., Cardoso, J.S., Horlings, H.M., Oliveira, S.P.: Gans vs. diffusion models for virtual staining with the her2match dataset. In: MICCAI Workshop on Deep Generative Models. pp. 120–130. Springer (2025)

[C] Attention-guided and MIL-constrained CycleGAN for high-fidelity virtual p16 and Ki-67 staining Lei et all, 2026.

Questions for authors

do you plan to make code and datasets public?

Reviewer_Kb1tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limited modeling of cell-level structures and the imperfect source–target correspondence commonly encountered in virtual staining. To this end, the authors propose CellDisc, which replaces the conventional binary discriminator with a cell segmentation-guided discriminator that explicitly incorporates cellular localization and instance-level geometry through cell probability maps and directional flow fields. Designed as a plug-and-play module, CellDisc can be integrated into a variety of paired GAN frameworks without modifying their generators. In addition, the authors establish three strongly paired, same-section datasets covering H&E-to-P63, H&E-to-P63&CK, and FFPE-to-H&E virtual staining tasks. Experimental results show that CellDisc improves structural fidelity across pixel-wise, perceptual, cell-level, and downstream classification evaluations.

Critical strengths

CellDisc introduces explicit cell-level structural supervision into GAN-based virtual staining while remaining compatible with multiple paired GAN architectures.

Critical weaknesses
  1. The authors constructed a strongly paired dataset, and CellDisc explicitly relies on the resulting cell-level correspondences for supervision. The experiments primarily demonstrate the advantages of cell-level supervision in the context of strong pairing. However, the strong coupling among dataset construction, method design, and the correspondence-dependent evaluation metrics implies certain limitations regarding the advantages demonstrated by CellDisc.
  2. The three datasets are presented as a major contribution of this work, yet the details regarding their construction are limited. The sources of the prostate cancer and breast cancer samples are also not clearly specified. In addition, although the datasets are split at the patient level, the total number of patients is not reported. Overall, given that the datasets are emphasized as an important contribution and that a substantial portion of the experiments is conducted on these newly constructed datasets, their detailed information should be clearly described.
Questions for authors
  1. The key contribution of CellDisc in Fig. 2 are mainly concentrated in (c) and (d). However, the green arrows preceding Fig. 2(c) should have clearly specified sources. In addition, Fig. 2(b) illustrates a limitation of existing methods, making its placement within the methodological framework somewhat abrupt.
  2. The authors claim that CellDisc can be plugged into paired GANs such as Pix2Pix, PyramidPix2Pix, and P2PDB. However, the manuscript does not clearly explain how CellDisc is integrated into conditional GANs.
  3. Regarding the dataset construction, each consisting of different staining conditions from the same tissue section. For the H&E-to-P63 and H&E-to-P63&CK datasets, a key question is how both H&E and IHC images were obtained from the same section and what staining sequence was used.
  4. For the FFPE-to-H&E task, the necessity of constructing a strongly paired dataset remain unclear. Since the manuscript does not clearly specify the imaging modality of the “FFPE” input images, it is difficult to determine how this dataset differs from existing paired unstained-to-H&E datasets. The authors should clarify the imaging modality of the FFPE and explain why such publicly available paired datasets were not used for external validation.
Reviewer_wiXJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CellDisc, a cell segmentation-guided discriminator for improving structural fidelity in virtual histological staining. Unlike conventional binary GAN discriminators, CellDisc predicts dense cell probability maps and directional flow fields, providing cellular localization and spatial organization supervision during adversarial training. The supervision is automatically derived from target images using CellposeSAM, and the proposed discriminator can be integrated into several paired GAN architectures without modifying their generators.

The authors additionally construct three strongly paired same-section datasets covering H&E-to-P63, H&E-to-P63&CK, and FFPE-to-H&E translation. Experiments demonstrate improvements in image quality, cell-level segmentation consistency, and stain-specific metrics across multiple GAN architectures. Comparisons with recent GAN- and diffusion-based methods, evaluations using independent cell segmentation models, and downstream cancer classification further support the effectiveness of the proposed framework.

Critical strengths
  1. The paper introduces a practical and well-motivated approach to incorporating cellular structure into adversarial virtual staining. By modifying the discriminator rather than the generator, CellDisc can be integrated into multiple existing GAN architectures. The experiments across different staining tasks and backbone generators provide meaningful evidence of its applicability and effectiveness.

  2. The experimental evaluation is extensive, and the construction of strongly paired same-section datasets is a valuable contribution. The study evaluates pixel-level, perceptual, cell-level, and stain-specific fidelity, while additional testing with independent segmentation models reduces concerns about relying exclusively on CellposeSAM-derived metrics. The downstream classification experiments also provide useful evidence of preserved diagnostic information.

Critical weaknesses
  1. The methodological novelty and the necessity of the proposed discriminator formulation require stronger justification. Incorporating segmentation-derived supervision into image generation is an established strategy, and CellDisc primarily relocates this supervision to a multitask discriminator. The current experiments do not directly compare against simpler alternatives, such as a conventional GAN discriminator combined with an auxiliary cell segmentation and flow-field consistency loss. Such comparisons are needed to establish whether the discriminator-centric design provides benefits beyond additional cellular supervision.

  2. The validation of biologically meaningful staining fidelity remains incomplete, particularly for H&E-to-IHC translation. Most quantitative results measure image similarity, cell segmentation consistency, or stain-region overlap, which do not necessarily establish correct cell-level biomarker expression. Additionally, the reported image-generation results lack uncertainty estimates across independent training runs or patients. Cell-level marker positivity agreement and patient-level statistical analyses would strengthen the biological and empirical reliability of the findings.

Questions for authors
  1. Have the authors compared CellDisc with a standard GAN discriminator supplemented by direct cell segmentation and flow-field consistency losses, using the same CellposeSAM pseudo-labels? Such a comparison would clarify whether the benefit arises specifically from the proposed discriminator formulation or from introducing additional structural supervision.

  2. Why are generated images assigned zero cell-probability targets but real-image flow-field targets during discriminator training? How does this asymmetric supervision affect the stability of the discriminator and the interaction between its probability and flow-field outputs?

  3. For the H&E-to-P63 and H&E-to-P63&CK tasks, have the authors evaluated cell-level biomarker expression agreement, such as positive-cell detection, P63 positivity, or CK localization consistency, using registered real IHC images or pathologist annotations? These results would help establish biological fidelity beyond morphological similarity.

  4. Are results from multiple independent training runs or patient-level bootstrap confidence intervals available for the primary generation metrics? In addition, could the authors report the numbers of independent patients and WSIs represented by each dataset to clarify the effective sample sizes?

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

Related papers

Discussion (0)

Sign in to comment.