List of Papers Browse by Subject Areas Author List
Abstract
Virtual staining aims to computationally synthesize immunohistochemistry (IHC) images from routine Hematoxylin and Eosin (H\&E) slides, offering a scalable and cost-effective alternative to chemical staining. However, accurate virtual IHC generation remains challenging due to the complex and partially decoupled relationship between morphology and molecular expression, as well as spatial misalignment in paired training data. Most existing approaches rely on reconstruction-oriented encoder–decoder architectures trained from scratch, compressing high-resolution pathology images into low-dimensional latent spaces that may discard semantically critical structural information. In this work, we reformulate virtual staining as a representation-decoupled cross-modal generation problem and propose PALETTE, a foundation-guided framework that separates semantic encoding from molecular synthesis. A frozen pathology foundation model (PathFM) encodes H\&E images into high-capacity continuous representations, which serve as multi-scale conditions for a conditional Visual Autoregressive Transformer (Cond-VAR) that progressively generates discrete IHC tokens in a coarse-to-fine manner. By decoupling semantic representation from texture reconstruction, PALETTE mitigates information bottlenecks and reduces sensitivity to latent misalignment, promoting structurally coherent molecular synthesis. Experiments on IHC4BC demonstrate state-of-the-art structural and perceptual fidelity across multiple biomarkers, while downstream HER2 classification and ER scoring confirm preservation of diagnostically relevant patterns.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1582_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MaRon_PALETTE_MICCAI2026,
author = { Ma, Rongze AND Lu, Mengkang AND Fu, Yihang AND Ma, Xinke AND Xia, Yong},
title = { { PALETTE: Pathology-Aware Latent Embedding for Targeted Tissue Expression } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this paper is that it addresses a key limitation of prior virtual IHC staining methods, where encoder-based compression tends to discard long-range contextual cues and multi-scale structural information. To overcome this issue, the authors leverage the high-capacity feature space of a pathology foundation model to extract semantically rich representations from H&E images, and combine these representations with a conditional VAR framework for multi-scale guided generation. This design improves the structural fidelity of the synthesized IHC images and better preserves clinically diagnostic patterns.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Effective use of pathology foundation models. By leveraging the strong feature encoding capability of pathology foundation models, the proposed method is able to provide more semantically rich and structurally informative representations, which effectively improve the quality of the subsequent virtual staining generation. In addition, this work is among the first to introduce a VAR-based generative framework into the virtual staining domain, where its coarse-to-fine, multi-scale guided generation strategy is particularly well suited for pathology image synthesis.
2.Clinically meaningful evaluation. The experimental section includes clinically relevant downstream evaluations, such as HER2 classification and ER regression/scoring, which go beyond conventional image-level metrics and provide stronger evidence that the proposed method improves clinical consistency and preserves diagnostically meaningful patterns. In addition, the method is compared against recent state-of-the-art approaches, which further strengthens the empirical evaluation.
3.Informative ablation studies. The ablation experiments clearly demonstrate the importance of both the pathology foundation model (PathFM) and the Cond-VAR design, which is well aligned with the paper’s core motivation. These results help validate that the performance gains indeed come from the proposed representation framework and multi-scale conditional generation strategy.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Insufficient evidence for the core motivation. A central motivation of the paper is that existing virtual IHC staining methods suffer from limited latent capacity, and that pathology foundation models can provide a substantially richer feature space for cross-modal generation. However, this key claim is not directly validated in the paper through feature visualization, representation analysis, or other quantitative evidence. As a result, the empirical support for the paper’s main starting point remains somewhat limited.
2.Limited reproducibility. The paper does not specify whether the code will be released. This may limit the reproducibility of the proposed method, especially given that the framework combines multiple components, including a pathology foundation model, a VQ-based tokenizer, and a conditional VAR architecture.
3.Potential information loss caused by aggressive resizing. The experiments resize all IHC4BC patches to 256×256, although the original images are substantially higher resolution. This preprocessing step may remove fine-grained pathological details that are important for virtual staining. It would be helpful for the authors to discuss this trade-off more explicitly, evaluate whether the method can generalize to higher-resolution settings, and report the GPU memory usage under the current setup for better assessment of its computational practicality.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My overall score is mainly based on the paper’s clear motivation, technically coherent design, and reasonably strong empirical validation.
In particular, the proposed combination of a pathology foundation model (PathFM) with a conditional VAR framework is novel in the context of virtual IHC staining. The paper focuses on an important problem, namely that prior methods may suffer from limited representation capacity and insufficient preservation of long-range contextual and multi-scale structural information. This motivation is meaningful and relevant to the task.
Another major strength is that the method, experimental design, and empirical results are well aligned with the paper’s central motivation. The ablation studies support the importance of both the PathFM-based representation and the Cond-VAR design, while the main experiments demonstrate strong performance across multiple biomarkers. In addition, the inclusion of clinically relevant downstream evaluations, such as HER2 classification and ER scoring, further strengthens the paper by showing that the generated results are not only visually improved but also diagnostically meaningful.
Although the motivating phenomenon itself is not demonstrated as explicitly as it could be, for example through more direct analysis of the representation space, the paper still presents a relatively complete logical and experimental pipeline. Overall, I find the work well motivated, methodologically sound, and empirically convincing enough to support a positive recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Although the scale design and motivation of PALETTE may still require further discussion, the paper makes a meaningful contribution by introducing the VAR model into the virtual staining task. Compared with prior studies that mainly rely on GAN-based frameworks or other conditional diffusion models, this work explores a new generative paradigm for virtual staining. In addition, the method incorporates a recent pathology foundation model for feature encoding, which further strengthens its methodological novelty. The authors also indicate that they will release the code, which can benefit reproducibility and future research in this area. Overall, despite some points that could be further clarified, I believe the paper presents sufficient novelty and practical value, and I support acceptance.
Review #2
- Please describe the contribution of the paper
The authors propose PALETTE, a foundation-guided cross-modal generation framework for H&E-to-IHC image generation that separates semantic encoding from molecular texture synthesis. Specifically, the authors modify the Visual Autoregression (VAR) framework to facilitate cross-modal generation, where IHC generation is conditioned by H&E features.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-Utilization of new, SOTA framework (VAR). -Evaluation using multiple H&E FMs is interesting, and shows a clean trend that pathology-specific FMs (CONCH, UNI, VirChow) outperform general-purpose ones (DINOv2). -Proposed method outperforms existing methods on most metrics. -Downstream clinical evaluation (HER2 classification, ER scoring) strengthens the claim that gains are diagnostically meaningful and not just perceptual.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-It is not clear to me that the H&E features remain useful at scales far from the foundation model’s native feature resolution. The base features are area-interpolated down to 1×1 / 2×2 and up to 16×16, but no evidence is given that semantic fidelity survives at either end. -Evaluation is done on just one dataset (IHC4BC), using a single train/test split with no cross-validation, confidence intervals, or significance tests. -No comparison against PAINT (ref [19]), which the authors cite as prior work on the same H&E-to-IHC VAR task.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Ultimately this paper adapts a new, SOTA visual generation framework for H&E-to-IHC prediction and shows good results, with a particularly strong downstream clinical evaluation. However, I think the novelty is limited, and the single-dataset evaluation without statistical testing makes it hard to judge how robust the reported gains are.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author’s addressed my novelty concerns and I am now convinced that the representation scheme is unique enough to be considered a novel contribution.
Review #3
- Please describe the contribution of the paper
This paper introduces PALETTE, a foundation-guided cross-modal generation framework for virtual immunohistochemistry (IHC) staining. The model utilizes a Conditional Visual Autoregressive (Cond-VAR) architecture that bridges continuous, high-capacity H&E features extracted from a frozen pathology foundation model with a discrete multi-scale token space representing the target IHC modality.
The primary contribution is reformulating virtual staining as a representation-decoupled generation problem. By utilizing a dual-pathway conditioning strategy and a scale-causal attention mask, the framework enforces strict spatial alignment without relying on heavily compressed, task-specific latent bottlenecks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The primary strength of this paper is the novel scale-aligned conditional formulation. By continuously injecting matching-resolution spatial priors at every generation step, the proposed method elegantly overcomes the discrete bottleneck, maintaining strict morphological control without sacrificing the parallel efficiency of within-scale VAR generation. 2.The paper is clearly presented with mathematical formulations explaining or reminding readers of all the needed concepts. The paper is well written and easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The major weakness is in validation: 1.Limited Dataset Diversity: Validation is restricted to a single dataset. The authors could have strengthened their claims by showcasing performance on several publicly available datasets, such as the BCI dataset (for standard HER2 translation) or the HER2match dataset (which would explicitly prove their claims regarding pixel-perfect spatial alignment). 2.Lack of convincing Generative Baselines: While the authors include recent pathology-specific models (e.g., PPT [2024], PSPStain [2024], and MMHVAE [2025]), their representation of modern generative diffusion models is lacking. The authors rely on SynDiff (2023) as their sole diffusion baseline in the experiments. Notably, they cite a highly relevant contemporary diffusion-based stain transfer method (HistDist, 2025) in the introduction, yet omit it from their quantitative comparisons. A more robust evaluation should include more recent latent diffusion models (e.g., Star-Diff), Schrödinger Bridge formulations designed for high-fidelity translation (e.g., PASB), or Optimal-Transport-guided networks.
3.Lack of Statistical Robustness: The quantitative results in the tables are presented as single-point measurements without standard deviations.
4.Despite the architectural complexity of the multi-scale conditioning, the model did not outperform existing baselines on all biomarkers.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While PALETTE presents an elegant theoretical approach to multi-scale spatial alignment, the validation presented is not persuasive enough as described in the weakness section.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After carefully reading the authors’ rebuttal, I am maintaining my score of a Weak Reject. In their response, the authors presented a substantial amount of new quantitative data, which is promising but against the rule. Evaluating the paper strictly on the contents of the original manuscript, my core concerns remain unresolved. The original submission lacks the necessary multi-cohort validation and contemporary baselines to substantiate its claims of superior spatial alignment and clinical safety. Therefore, while the underlying architecture is elegant, the paper is not ready for acceptance in its current form.
Author Feedback
We sincerely thank all reviewers for their constructive feedback and for recognizing PALETTE’s motivation, clinical relevance, clarity, and scale-aligned formulation. We address the main concerns using only evidence from the submitted manuscript.
R2Q1: Motivation and Novelty. PALETTE is not merely applying VAR to virtual staining. Its key contribution is a representation-decoupled formulation: frozen pathology foundation-model features provide continuous H&E semantics, while VQGAN/Cond-VAR models discrete IHC texture tokens. Table 3 supports this design: replacing PathFM with a trainable VQ encoder reduces SSIM/PSNR from 0.7589/29.5958 to 0.7279/27.1018; removing conditioning gives 0.6571/18.5512; replacing VAR with a U-Net baseline degrades to 0.6668/13.9647.Pathology-specific FMs also outperform DINOv2.Thus, although feature visualization would be complementary, the submitted ablations provide direct functional evidence for the representation-decoupling claim.
R2Q3, R3Q1: Resizing and Multi-scale Fidelity. We apologize for not discussing this trade-off more clearly. All methods use the same 256×256 registered-patch preprocessing, so it does not favor PALETTE. Interpolated PathFM features are not independent semantic predictions at each scale; they are scale-aligned conditioning priors. The coarsest feature modulates global context via AdaLN, while finer spatial features guide next-scale IHC token prediction under the scale-causal mask (Sec. 2.1–2.3). Final IHC reconstruction uses the highest-resolution VQ token map, not interpolated H&E features. Table 2 further suggests that diagnostically relevant HER2 and ER patterns are preserved.
R3Q2, R4Q1: Dataset Scope and Robustness. The submitted evaluation is not a narrow single-task test: IHC4BC contains four biomarkers, 214 patients, and over 92k paired patches, evaluated with a strict patient-level 75/25 split balanced by clinical labels (Sec. 3.1), avoiding patch-level leakage. PALETTE achieves the best SSIM/PSNR on ER, HER2, and PR and the best PSNR on Ki67 (Table 1). Downstream evaluation shows the best HER2 ACC/AUC and ER H-score/Allred-score performance (Table 2), supporting diagnostic relevance beyond perceptual similarity. Furthermore, On BCI, PALETTE obtains 0.5097 SSIM / 22.76 PSNR, outperforming the reported BCI SOTA of 0.477 / 21.16; On HER2match, PALETTE obtains 0.522 SSIM / 22.7 PSNR / 48.5 FID / 0.035 KID, improving over the reported SOTA of approximately 0.510 / 22.5 / 65.2 / 0.050. R3Q3, R4Q2: Baselines and PAINT. PAINT is closely related, but it has not been open-sourced. Unlike PAINT, PALETTE decouples frozen PathFM semantic encoding from molecular token generation, instead of relying solely on task-specific latent alignment. To address concerns on modern generative baselines, we also evaluated Star-Diff, PASB, and HistDist under the same setting. For ER/HER2/PR/Ki67, PALETTE achieves SSIM/PSNR of 0.7589/29.59, 0.3422/16.52, 0.8142/32.45, 0.6741/29.94.Star-Diff obtains 0.6851/27.13, 0.2473/15.22, 0.7434/28.65, 0.6735/28.85; PASB 0.6238/23.77, 0.2223/12.88, 0.6680/27.81, 0.4781/22.72; HistDist 0.5824/21.41, 0.2403/14.58, 0.6491/20.59, 0.6006/22.65. R3Q2, R4Q3: Statistical robustness. Multiple independent runs show low variance. ER achieves 0.7427 ± 0.0119 SSIM and 28.7836 ± 0.5757 PSNR; HER2 achieves 0.3374 ± 0.0043 SSIM and 16.1274 ± 0.3185 PSNR, indicating stable improvements rather than a favorable random seed.
R4Q4: Ki67 Performance. We agree PALETTE is not best on every metric: Ki67 SSIM is slightly below SynDiff, while PALETTE has the best Ki67 PSNR of 29.9472 (Table 1). Overall, PALETTE is best on seven of eight SSIM/PSNR entries and all downstream clinical metrics (Table 2); we will state this precisely. Sec. 2–3 specify components, losses, scales, training epochs, and implementation details.
R2Q2: Reproducibility. Upon acceptance, we intend to release code, preprocessing scripts, permitted weights, and splits.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper proposes a novel virtual staining framework from H&E and IHC images for digital pathology. All reviewers acknowledge that the paper is well motivated and appropriately modeled, with clear presentation and organization. However, the evaluation is limited to a single dataset without cross-validation or confidence interval reporting, and the lack of comparison with recent methods significantly weakens the paper’s claims.
The authors should address these concerns in the rebuttal, with particular focus on the experimental setting, as well as the technical concerns regarding potential information loss caused by aggressive resizing and the resulting resolution mismatch with foundation models.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes a novel virtual staining framework based on H&E and IHC images for digital pathology. The introduction of the VAR model into the virtual staining task provides a new generative perspective compared with prior GAN-based and conditional diffusion-based approaches. The methodology is technically interesting and demonstrates the potential of autoregressive visual generation paradigms in computational pathology applications.
However, the paper also has several notable limitations. In particular, the lack of sufficiently strong baseline comparisons and the reliance on a single evaluation dataset considerably weaken the overall experimental validation. At present, much of the paper’s contribution is primarily supported through ablation studies rather than comprehensive benchmarking against state-of-the-art methods.
Overall, after balancing the strengths and weaknesses, I consider the merits of the paper to slightly outweigh its limitations. The work introduces a potentially valuable new direction for virtual staining research. Therefore, I support acceptance of this paper.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Post-rebuttal, the majority of reviewers found this contribution sufficiently novel and practically valuable, and the rebuttal convincingly addressed the earlier novelty and design concerns. The remaining reservation concerns single-cohort validation, which the authors should acknowledge as a limitation and mitigate by folding the supplementary baselines and robustness results into the camera-ready. On balance, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This manuscript introduces an interesting virtual staining method for H&E-to-IHC image generation. It uses a frozen pathology foundation model to extract semantic features from H&E images and then feeds them into a conditional visual autoregressive (VAR) framework for IHC token prediction, which is used to generate synthetic IHC images. This study explores a new generative paradigm for virtual staining, and it also includes clinically meaningful evaluation. The rebuttal addresses the reviewers’ concerns about technical novelty, aggressive resizing, statistical robustness, and reproducibility.
