List of Papers Browse by Subject Areas Author List
Abstract
Although deep learning has advanced Whole Slide Image (WSI) Analysis, tissue artifacts like bubbles and folds often cause silent failures by concealing essential morphology. Current pathology image restoration methods are mostly restricted to small patches, struggling to maintain global structural coherence at a megapixel scale. We in troduce RestorePath, a framework for globally consistent megapixel scale inpainting that reconstructs diagnostic structures in histological image to prevent incorrect high-confidence predictions and lower error rates. Our model utilizes a Latent Diffusion Model (LDM) conditioned on Pathology Foundation Model (PFM) embeddings, integrating Large Kernel Attention (LKA) to manage long-range dependencies during random patch diffusion. Enhanced by Distance-Weighted Interpolation (DWI) and an Adaptive Guidance Scale (AGS), RestorePath ensures structural consistency and fidelity by modulating information from surrounding patches. Evaluations across TCGA-BRCA, BACH, and Camelyon16 datasets for images ranging from 512 to 4608 pixels demonstrate state-of-the-art performance in maintaining histological consistency. RestorePath significantly improves downstream Computational Pathology (CP) tasks, outperforming both raw artifact images and the conventional Detect-and-Discard (D&D) approach. The code is available at https://github.com/PathfinderLab/RestorePath.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6076_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/PathfinderLab/RestorePath
Link to the Dataset(s)
TCGA BRCA dataset: https://portal.gdc.cancer.gov/projects/TCGA-BRCA
Camelyon16 dataset: https://camelyon16.grand-challenge.org/Data/
BACH dataset: https://iciar2018-challenge.grand-challenge.org/Dataset/
BibTex
@InProceedings{LeeHye_PatchtoGlobal_MICCAI2026,
author = { Lee, Hyeseong AND Kim, Eunsu AND Bappy, Dawen M. AND Kim, Ho Heon AND Lee, Youngsuk AND Chun, Se Young AND Choi, Jang-Hwan AND Lee, Sung Hak AND Ahn, Sangjeong},
title = { { Patch-to-Global: Random Patch Diffusion for Globally Consistent Megapixel Artifact Inpainting in Whole Slide Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents an inpainting approach, which is particularly designed to compensate large (megapixel) missing information in histological whole slide images. This paper presents an apparently strong and well-executed approach with impressive empirical results. Besides qualitative evaluation based on simulated degradations, they also investigate the impact on downstream classification tasks. Particularly the latter is highly interesting for real world application.
- 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.
- the paper addresses a relevant and timely problem, with a clearly defined experimental protocol and a seemingly careful evaluation setup
- quantitative results are mostly strong across multiple benchmarks and settings, suggesting that the proposed approach is effective under the tested conditions
- the ablation studies are reasonably structured and aim to isolate the impact of key design choices
- the method appears technically sound and well-motivated, and the paper itself is clearly 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 reported performance gains appear unusually large and potentially over-optimistic, raising concerns about how well these results would translate to realistic downstream settings.
- while the evaluation methodology seems correct, the magnitude of the gains makes it difficult to assess external validity
- the “dd” scenario is extremely degraded, significantly more so than what would typically be expected in practical settings. This raises concerns that the scenario may not be representative or may exaggerate failure modes
- the downstream task has only been evaluated with the proposed method and not with the sota appraoches. In order to have a more realistic baseline, it would be extremely interesting to see how the other appraoches perform here.
- details on the training of the benchmark appraoches are missing. Where they individually trained on histological data?
- 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 provide sufficient information for 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?
- overall interesting study, particularly the downstream task evaluation is highly relevant
- however, a comparison with a state-of-the-art method is missing in the downstream evaluation
- the quantitative evaluation is hard to assess - also due to partly inconsistent results
- 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 rebuttal addresses my main concerns by clarifying the baseline training protocol and strengthening the clinical validity argument and providing additional evidence that the observed performance gains are not solely due to unrealistic degradation settings. While some questions remain regarding the absence of explicit downstream SOTA comparisons and the external validity of certain metrics, the responses sufficiently improve the overall credibility of the experimental claims. Given the relevance of the problem and the generally strong empirical results, I am moving my decision to a weak accept.
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is the proposal of RestorePath, a patch-to-global diffusion framework for globally consistent megapixel-scale artifact restoration in whole slide images (WSIs). Instead of only repairing small local patches, the method extends patch-level inpainting to large pathological regions by combining a PFM-conditioned latent diffusion model, Large Kernel Attention (LKA) for long-range dependency modeling, and a Random Patch Diffusion inference scheme with Distance-Weighted Interpolation (DWI) and Adaptive Guidance Scale (AGS) to preserve both local fidelity and global histological consistency. The paper further shows that this restoration is not only visually better, but also improves downstream computational pathology tasks compared with using corrupted images directly or applying detect-and-discard quality control.
- 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.
First, it goes beyond prior pathology artifact restoration methods that mainly operate on small local patches, and instead explicitly formulates the problem as globally consistent megapixel-scale restoration. This is a meaningful contribution because the paper clearly identifies that large artifacts in WSIs cannot be adequately handled by isolated patch-level inpainting and therefore require a patch-to-global solution. Second, the proposed framework is technically coherent and well matched to this problem setting. The combination of a PFM-conditioned inpainting latent diffusion model, Large Kernel Attention (LKA) for modeling long-range dependencies, and Random Patch Diffusion for extending patch-trained restoration to large corrupted regions is logically designed, and the additional DWI and AGS modules are also well motivated within this framework. Third, the paper evaluates the method with pathology-aware metrics rather than relying only on generic image similarity measures. In particular, the use of CONCH-based FID and embedding similarity makes the evaluation more relevant to pathological feature preservation and histological consistency, which is more appropriate for this domain than standard natural-image restoration benchmarks alone.
- 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, The first concern is about clinical validity. The paper frames artifact restoration as preferable to detect-and-discard because discarding may remove sparse lesions in “needle-in-a-haystack” settings. However, the proposed method does not show that the hidden content under the artifact can be faithfully recovered, especially when that hidden content is tumor or other diagnostically decisive morphology. The method explicitly reconstructs masked regions by conditioning on embeddings from surrounding artifact-free tissue and propagating nearby context into the corrupted region via Distance Weighted Interpolation (DWI). In other words, when the target patch is heavily or fully occluded, the model is encouraged to infer what is plausible given nearby tissue, not what was actually present beneath the artifact. Though it argues against detect-and-discard, but does not address the reason that such strategies exist in medical workflows: artifact regions are often excluded because synthesized replacements may be unsafe.
2, The methodological novelty is limited and several modules appear engineering-driven. The overall pipeline is still an inpainting latent diffusion model with pathology embeddings, plus a large-image inference strategy. Similar directions already exist in prior pathology restoration and diffusion-based large-image generation, including Artifusion[6], ArtiDiffuser[21], DiffInfinite[1], LRDM[5], ZoomLDM[25] and even LatentArtifusion [A]. The proposed additions—LKA, distance-weighted interpolation, and adaptive guidance scaling—are reasonable, but they read more like task-specific engineering choices than a clearly new formulation. The paper would need stronger analysis to argue that these are conceptual advances rather than practical tweaks.
3, The baseline comparison is limited to general image inpainting methods. The main quantitative comparison is against LaMa, CoordFill, and SDM, which are mostly general-image inpainting baselines. However, the paper itself discusses Artifusion and ArtiDiffuser as prior pathology-specific artifact restoration methods, yet does not include them in the main comparison. This weakens the empirical claim substantially, because the paper’s main novelty is supposed to be improving medical/pathology restoration beyond patch-level methods. A natural question is whether these prior medical methods, combined with the same large-image inference wrapper, would perform similarly.
4, The method is relatively heavy and not clearly practical for deployment. The paper states that separate models are trained for each dataset, which raises concerns about scalability and practical use. This is especially important because some general methods already achieve competitive performance on certain metrics, such as LaMa on Camelyon16 and BACH LPIPS. The paper does not provide computational costs for training and inference. As a result, the method appears effective but operationally heavy.
5, The evaluation depends heavily on synthetic artifacts. A substantial part of the validation and testing protocol is based on synthetically generated artifacts, including TCGA-BRCA and Camelyon16 test settings. This makes the experiments more controlled, but also less convincing for real clinical use, since real pathology artifacts can be far more irregular and heterogeneous. The paper’s claim is about restoring real WSI artifacts, yet it does not provide enough evidence on genuine real-world corrupted slides. Thus, the experimental setting may overestimate practical effectiveness.
6, The DWI ablation behavior is not sufficiently explained. In Table 1, removing DWI causes a large drop on TCGA-BRCA, but only limited changes on Camelyon16 and BACH. The paper gives a brief explanation that TCGA-BRCA contains more opaque artifacts, while bubbles in the other datasets preserve local textures. However, this discussion is too shallow for one of the paper’s main proposed modules. A stronger paper should analyze more clearly when DWI is necessary and how its benefit depends on artifact type and visibility. Also, since DWI brings only marginal gains on Camelyon16 and BACH, where bubble artifacts dominate, does this suggest that DWI may be unnecessary in bubble-heavy settings? If so, it would be important to discuss whether DWI could be omitted in such cases to improve efficiency, especially since it may introduce additional computational overhead, which is not analyzed in the paper.
Minor:
- In conclusion, “we employs” should be “we employ”
- Table 3 cites ResNet50 [8], but reference [8] is the DenseNet paper rather than the ResNet paper.
[A] LatentArtiFusion: An Effective and Efficient Histological Artifacts Restoration Framework
- Please rate the clarity and organization of this paper
Satisfactory
- 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 provide sufficient information for 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please see the weakness part.
- Reviewer confidence
Very confident (4)
- [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 reading the rebuttal, I maintain my reject recommendation. In my view, the paper still falls below the MICCAI acceptance bar. The rebuttal does not sufficiently resolve my main concerns regarding clinical validity, baseline comparison, novelty, and experimental realism.
First, the response on clinical validity and safety does not address my central concern. My point was not whether restored images can improve downstream AUC relative to Artifact or D&D, but whether the restored content is faithful to the hidden pathology rather than merely plausible and classifier-friendly.
Second, I remain unconvinced by the rebuttal’s argument on baseline comparison. The lack of comparison with medical/pathology-specific restoration baselines is still a significant weakness. Given the obvious domain gap between natural images and pathology images, it is entirely within expectation that general-image inpainting methods transfer imperfectly to this setting. This does not by itself validate the proposed method. More importantly, I do not find the claim that prior pathology restoration methods are “architecturally unable to ingest megapixel artifacts” persuasive. The most straightforward adaptation would be to process megapixel corrupted regions in a patch-wise manner and then evaluate how far such medical baselines can go under the same large-image setting. Even if imperfect, that would still be a much more relevant comparison than relying primarily on natural-image baselines.
Third, I continue to find the novelty limited. The rebuttal does not change my impression that the main components are largely engineering-driven.
Finally, the paper still relies heavily on synthetic artifacts, which weakens the realism of the evaluation. This is not a minor detail: for example, the TCGA-BRCA validation regions are corrupted with synthetic artifacts generated via failure modes, the Camelyon16 test regions are masked with synthetic bubbles, and the BACH test setting is also based on bubble corruption.
Additionally, the computational costs (~700s) is TOO much in my view, and it is around 5minutes more compared with SDM. I do not consider this overhead acceptable.
Review #3
- Please describe the contribution of the paper
This paper addresses the problem of large-scale artifact restoration in whole slide images and aims to improve restoration quality at the megapixel level while preserving global histological consistency. The work is further strengthened by evaluating not only image restoration quality but also the impact on downstream computational pathology tasks.
- 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 paper addresses a clinically relevant and practically important problem. Large-scale artifacts in WSIs can indeed affect downstream diagnosis and model reliability, and the paper clearly motivates why simple detect-and-discard strategies may be suboptimal in pathology settings. 2.The overall framework is reasonably well designed for the target problem. The combination of LKA for long-range dependency modeling and DWI/AGS for high-resolution inference is intuitive and matches the goal of preserving global histological consistency during large-region restoration. 3.The experimental section is relatively comprehensive. The downstream-task evaluation improves the practical relevance of the work.
- 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.Some empirical gains are modest. For example, the proposed method is not consistently the best on LPIPS, and some improvements over the ablated variants are relatively small, which weakens the evidence that each proposed component yields a strong practical advantage. 2.Although the task is motivated by real WSI artifacts, most experiments are conducted under synthetic masking or simulated artifact settings, making it unclear how well the method generalizes to real-world artifact distributions and clinical scenarios. 3.The paper emphasizes “megapixel-scale” restoration, but the analysis specifically supporting this claim is still limited.
- Please rate the clarity and organization of this paper
Satisfactory
- 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?
I recommend a weak accept. The paper addresses a clinically relevant problem and presents a reasonably well-executed framework with useful downstream validation, although the experiments are somewhat limited for a stronger recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all reviewers for their constructive feedback. We address the detailed comments below.
1.Clinical Validity & Safety [R2-C1]. R2 argues synthesizing occluded patches is clinically unsafe, inferring plausible tissue instead of recovering hidden tumors. Table 2 directly tests this risk by overlaying bubbles on 38 Camelyon16 tumor regions. Restored AUC matches Origin within one std-dev (ABMIL 90.3 vs 89.6, TransMIL 85.0 vs 86.2, CLAM 90.3 vs 88.5). If the model merely hallucinated normal tissue, AUC would have collapsed to D&D levels (70.8/59.4/71.6). This AUC recovery and Fig 4’s attention relocalization prove diagnostic morphology is authentically preserved, preventing the unsafe silent failures R2 assumes.
2.Pathology-Specific Baselines & Novelty [R2-C2, R2-C3] Patch-level models like Artifusion[6], ArtiDiffuser[21], and LatentArtiFusion[A] lack global conditioning and are architecturally unable to ingest megapixel artifacts. Our SDM+DiffInfinite-wrapper baseline faithfully represents this class. Table 1 shows it collapses on fully occluded patches (FID_C 25-34% worse) and Fig 2 visually confirms inconsistency. LRDM[5] and ZoomLDM[25] are generators, not conditional inpainters. RestorePath is the first principled framework extending patch-trained diffusion to globally consistent megapixel inpainting via DWI and AGS, marking a structural contribution.
3.Downstream Comparison & Baseline Training [R1-C4, R1-C5] All baselines (LaMa/CoordFill/SDM) were fine-tuned on the same 256x256 histology patches using their official codebases and masks. Downstream SOTA comparison is logically self-defeating: on the two downstream datasets, LaMa’s FID_C is 1.9-3.1x worse (Camelyon16 83.28 vs 44.57, BACH 135.61 vs 44.37) and Fig 2 confirms morphological destruction. Feeding morphologically degraded patches into MIL/CNN is a predicted failure adding no information beyond FID_C. We benchmarked Origin/Artifact/D&D as the clinically meaningful alternatives.
4.External Validity, D&D Fairness, Real-World Generalization [R1-C1/C2/C3, R2-C5, R4-C2] The Restored-vs-D&D gap reflects clinical reality. Standard QC pipelines strictly exclude artifacts, which our D&D simulates. As shown in Fig 1(a), real megapixel artifacts span 44-71% of the affected region, making patch-level soft handling infeasible. Paired quantitative evaluation on real artifacts is mathematically impossible (real artifacts destroy ground truth), so we (i) train with 50% real GrandQC masks, (ii) apply failure-mode synthetics [22] matching real distributions, and (iii) test downstream on real WSIs using controlled overlays to preserve clinically authentic tissue heterogeneity and lesion distributions.
5.Megapixel Scope, DWI Ablation, LPIPS, Marginal Gains [R4-C1/C3, R2-C6] Megapixel artifacts comprise 73% (TCGA-BRCA), 100% (BACH), and 57% (Camelyon16) of our synthetic cases. DWI’s dataset-dependent gain is by design, activating only when in-patch context is absent. TCGA-BRCA markings completely occlude tissue, making DWI essential since no original signal remains (FID_C 110.48 to 29.66). Conversely, sparse Camelyon16/BACH bubbles leave residual textures where cross-attention suffices. These marginal gains reflect saturation, not module weakness. While LPIPS favors LaMa’s texture matching, FID_C and Emb Sim measure true pathological preservation where RestorePath remains consistently SOTA.
6.Compute [R2-C4]. For 2048x2048 inference (RTX A6000), LaMa (~0.4s) and CoordFill (~0.1s) are fast but visually destroy morphology (FID_C 4-5x worse, Table 1). Among diffusion models capable of global consistency, RestorePath takes ~760s vs SDM’s ~400s. This 1.9x overhead for megapixel coherence is strictly justified against the cost of physical re-scanning (hours-days) and silent diagnostic failures.
7.Minor [R2] Typo and citation will be corrected. Code will be released upon acceptance.
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.
The paper proposes an inpainting method for tissue artifacts in whole slide images based on a latent diffusion model. It aims to provide better performance for coherent inpainting at a megapixel scale with improved performance on downstream tasks.
Reviewers appreciate the general setup of the evaluation, strong results and the insights derived by ablation studies. They further highlight the attractiveness of an inpainting approach that addresses globally consistent restoration, reducing the risks of new artifacts being introduced by the inpainting model itself. The overall setup is further seen as convincingly designed.
However, reviewers also raise important concerns. First, performance gains appear in part potentially over-optimistic, raising doubts on generalizability. Related to this, there is the concern that inpainting does not address whether hidden content (e.g., sparse lesions) can be faithfully reconstructed. Second, most investigations are conducted on synthetic data and specific degradation scenarios are deemed unrealistic beyond what would typically be seen in clinical practice. Third, the downstream task was not evaluated on competitive state-of-the-art approaches, again raising doubts on the net benefit of the proposed approach.
In their rebuttal, authors should address concerns regarding the level of conceptual advances in the proposed approach. Additionally, they should address concerns regarding computational complexity / runtimes for the proposed method. Additionally, it should be discussed to what extent performance, especially benefit measured here on strong degradations, can be representative for generalization to real-world scenarios. Reviewer R#2 further mentioned concerns regarding the ablation study, which should be further addressed.
- 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.
After rebuttal, reviewer rating continues to be mixed. Some concerns regarding baseline training and clinical validity were resolved for R#1 but not for R#2.I share concerns regarding additional, more powerful baselines with R#2 and am missing an integration of compared approaches (e.g., SDM) in the downstream task analysis in Table 2.At the same time, I agree that the purpose of such an inpainting task - if used in the manner evaluated in the paper itself - cannot be to recover lost information but to restore downstream performance. Given the overall assessment, I lean towards a weak accept recommendation.
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.
This paper addresses an important and clinically relevant problem: restoring large-scale artifacts in whole-slide pathology images while preserving global histological consistency. Reviewers generally appreciated the empirical evaluation, pathology-aware assessment metrics, and inclusion of downstream computational pathology tasks.
Several concerns were raised regarding clinical validity, methodological novelty, computational cost, reliance on synthetic artifacts, and absence of comparisons with pathology-specific restoration methods. The rebuttal provided additional clarification regarding baseline training, rationale for selected comparisons, role of synthetic artifacts, and clinical motivation. These responses satisfactorily addressed a number of concerns. While questions remain regarding real-world generalisation and computational efficiency, the overall contribution, experimental quality, and potential impact put it marginally above the acceptance bar.
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.
The paper proposes a latent diffusion model for megapixel-scale artifact inpainting in whole slide images. While initial reviews raised concerns regarding computational efficiency, synthetic degradation settings, and baseline comparisons, the authors’ rebuttal and the subsequent change in stance from Reviewer #1 effectively mitigated these concerns. The authors clarified that the baseline training protocols were equivalent, strengthened the clinical validity argument through downstream evaluation, and provided evidence that the performance gains are not solely artifacts of unrealistic degradation settings. Given the relevance of the problem—addressing the limitations of “detect-and-discard” strategies—and the demonstrated ability to maintain histological consistency at the megapixel scale, the paper is now recommended for acceptance.