List of Papers Browse by Subject Areas Author List
Abstract
Deep learning advances have revolutionized automated digital pathology analysis. However, variations in staining protocols and imaging conditions can introduce significant color variability. In deep learning, such color inconsistency often reduces performance when deploying models on data acquired under different conditions from the training data, a challenge known as domain shift. Many existing methods attempt to address this problem via color normalization but suffer from several notable drawbacks such as introducing artifacts or requiring careful choice of a template image for stain mapping. To address these limitations, we propose a trainable color normalization model that can be integrated with any backbone network for downstream tasks such as object detection and classification. Based on the physics of the imaging process via the Beer-Lambert law, our model architecture is derived via algorithmic unrolling of a nonnegative matrix factorization (NMF) model to extract stain-invariant structural information from the original pathology images, which serves as input for further processing. Experimentally, we evaluate the method on publicly available pathology datasets and an internally curated collection of malaria blood smears for cross-domain object detection and classification, where our method outperforms many state-of-the-art stain normalization methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4658_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xutianyue/BeerLaNet
Link to the Dataset(s)
Malaria Detection Dataset: https://zenodo.org/records/8358829
Whole Blood Cell Detection: https://github.com/Shenggan/BCCD_Dataset
Breast Cancer Classification: https://wilds.stanford.edu/datasets/#camelyon17
BibTex
@InProceedings{XuTia_Adaptive_MICCAI2025,
author = { Xu, Tianyue AND Wu, Yanlin AND Tripathi, Abhai K. AND Ippolito, Matthew M. AND Haeffele, Benjamin D.},
title = { { Adaptive Stain Normalization for Cross-Domain Medical Histology } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15966},
month = {September},
page = {24 -- 33}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposed a new stain normalization model via algorithmic unrolling to solve non-negative matrix factorization (NMF). The original stain normalization through NMF needs to choose a template image and corresponding color spectra matrix, carefully select sparse regularization strength lambda, and the rank of factorization r. But the proposed BeerLaNet alleviates the problems by end-to-end training to learn parameters in downstream tasks. The experiments show advantages of BeerLaNet in some object detection and image classification 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.
The proposed method in the paper alleviates the problems of choosing template image and hyper-parameters selection in original NMF model by integrating optimization with downstream tasks to apply end-to-end training.
The paper demonstrated various types of downstream tasks, including object detection and image classification, to evaluate the effectiveness of proposed method, and compared with other existing stain normalization methods.
The results show that the proposed method performs best in some object detection and image classification tasks.
- 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 proposed method generates r channels for each image, but the optimization method does not fix the focus or meaning for each channel. And the focus/meaning for each stain density channel may vary on different images, which theoretically causes challenges for (detection, classification) encoders that take those channels as inputs in the pipeline.
The paper states that the gamma and lambda is updated via supervision on downstream tasks, but there is no such clue in the full details provided in Algorithm 1.
The r-channel image is passed through a 1x1 convolution layer to map the image back to 3-channel image. But the 3-channel image are probably no longer RGB channel image as other stain normalization methods. The normalization is task specific and not generic.
The original stain normalization method that needs template image can actually also introduce adaptive new color spectra matrix and perform end-to-end training to get optimum new color spectra matrix. The paper didn’t compare with this setting to have a fairer comparison.
There is no ablation study of colored component r. The paper states that the method allows r to be larger than expected number of components, but further ablation study is missing to demonstrate that larger r (like 8) does not harm the experiment results.
The paper didn’t mention time complexity comparisons. The proposed method is implemented with downstream task at the same time, so it takes more time and iterations than other stain normalization that only needs to be done once in the pre-processing.
The proposed method does not show constant advantages over other stain normalization methods in experiments.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/en/REVIEWER-GUIDELINES.html
The statement “the use of PCA for estimating S and D has clear deficiencies, as it requires the columns of the matrices to be orthogonal, which is clearly not realistic for the problem.” in the paper is not correct. Based on Macenko’s paper [1], the calculated 3 orthogonal vectors are for finding a 2D plane in 3D space by eliminating one vector, and are not directly used for S. The final color spectra of different stains are mapped to the 2D plane and are not necessarily orthogonal to each other.
[1] Macenko, Marc, et al. “A method for normalizing histology slides for quantitative analysis.” 2009 IEEE international symposium on biomedical imaging: from nano to macro. IEEE, 2009.
- 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?
Even though the paper proposed BeerLaNet that alleviates the problems of template image choosing and selection of parameters in original stain normalization by end-to-end training to learn parameters in downstream tasks, the method theoretically causes challenges during training. Also, the proposed method is not generic and takes more time to train. Additionally, the experiment results show the BeerLaNet does not always have advantages, and ablation studies on the colored component r to confirm the claim of flexible choice of r is missing.
- 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 checking the rebuttal, most of my concerns still exist. The details of my further comments on each concern after rebuttal can be found below:
-The proposed method generates r channels for each image, but the optimization method does not fix the focus or meaning for each channel. And the focus/meaning for each stain density channel may vary on different images, which theoretically causes challenges for (detection, classification) encoders that take those channels as inputs in the pipeline.
After rebuttal: The authors stated that their initialization is fixed to maintain consistency of each channel, but initialization value can be easily changed during training. The theoretical weakness still exists.
-The paper states that the gamma and lambda is updated via supervision on downstream tasks, but there is no such clue in the full details provided in Algorithm 1.
After rebuttal: How the gamma and lambda is updated is still unclear.
-The r-channel image is passed through a 1x1 convolution layer to map the image back to 3-channel image. But the 3-channel image are probably no longer RGB channel image as other stain normalization methods. The normalization is task specific and not generic.
After rebuttal: This weakness still exists.
-The original stain normalization method that needs template image can actually also introduce adaptive new color spectra matrix and perform end-to-end training to get optimum new color spectra matrix. The paper didn’t compare with this setting to have a fairer comparison.
After rebuttal: not solved, the paper still misses a fair comparison with other stain normalization method.
-There is no ablation study of colored component r. The paper states that the method allows r to be larger than expected number of components, but further ablation study is missing to demonstrate that larger r (like 8) does not harm the experiment results.
After rebuttal: the authors admitted the missing reports of this ablation study.
-The paper didn’t mention time complexity comparisons. The proposed method is implemented with downstream task at the same time, so it takes more time and iterations than other stain normalization that only needs to be done once in the pre-processing.
After rebuttal: the quantitative time complexity is not provided
-The proposed method does not show constant advantages over other stain normalization methods in experiments.
After rebuttal: Based on provided experiment results, the proposed method is generally the best on half of the datasets.
-The statement “the use of PCA for estimating S and D has clear deficiencies, as it requires the columns of the matrices to be orthogonal, which is clearly not realistic for the problem.” in the paper is not correct. Based on Macenko’s paper [1], the calculated orthogonal vectors are for finding a 2D plane in 3D space, and are not directly used for S, the final color spectra of different stains are not necessarily orthogonal to each other.
[1] Macenko, Marc, et al. “A method for normalizing histology slides for quantitative analysis.” 2009 IEEE international symposium on biomedical imaging: from nano to macro. IEEE, 2009.
After rebuttal: This comment is not responded by authors.
Review #2
- Please describe the contribution of the paper
The paper presents a trainable color normalization model, invariant to backbones and downstream tasks for digital pathology slides. The method is called Beer-LaNet is built upon non-negative matrix factorization and algorithmic unrolling.
- 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 is easy to follow and the different components seem to boost the performance of the method.
- The topic and the problem that the authors try to address in this study is very interesting.
- 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.
- Currently, different very powerful encoders and foundation models (for example UNI [1], CONCH[2], Phikon[3], DinoBloom [4]) have been proposed in the literature. These models are able to provide some very good and invariant representations of tiles. I find that the current paper does not really discuss about them and compare with them.
- Continuing my previous comment, I find the baselines used (e.g. ResNet18) a bit weak making difficult to highlight the advantages of the method.
- An ablation study to benchmark the different components of the method is missing.
[1] Chen, Richard J., et al. “Towards a general-purpose foundation model for computational pathology.” Nature Medicine 30.3 (2024): 850-862. [2] Lu, Ming Y., et al. “A visual-language foundation model for computational pathology.” Nature Medicine 30.3 (2024): 863-874. [3] Filiot, Alexandre, et al. “Phikon-v2, a large and public feature extractor for biomarker prediction.” arXiv preprint arXiv:2409.09173 (2024). [4] Koch, Valentin, et al. “DinoBloom: a foundation model for generalizable cell embeddings in hematology.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024.
- 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.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is mainly based on the applicability of the method in an era of foundation models. Some better benchmark of the method with more recent models would highlight better the soundnes of the method.
- Reviewer confidence
Very confident (4)
- [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
Review #3
- Please describe the contribution of the paper
The paper proposes BeerLaNet, an adaptive stain normalization method designed to improve deep learning performance in histopathology by addressing stain variability issues. BeerLaNet utilizes the Beer-Lambert law and leverages an algorithmic unrolling of Nonnegative Matrix Factorization (NMF), leading to a physics-informed, end-to-end trainable normalization model.
- 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 proposed method integrates classical physics-based models (Beer-Lambert law, NMF) with modern deep-learning paradigms via algorithmic unrolling.
- The framework can be seamlessly combined with various downstream backbone networks.
- Experimentally demonstrates strong performance and generalization capability across multiple challenging datasets.
- 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.
- While the approach is practically effective, the underlying idea (algorithmic unrolling of Beer-Lambert law and NMF) is incremental rather than fundamentally novel.
- The extent to which the model goes beyond simple unrolling and integration into existing frameworks could be clearer.
- Could benefit from more rigorous comparisons with other deep unrolling methods.
- Since the proposed method belongs to stain normalisation, it is beneficial to provide the comparison of stain normalised results.
- The code is not provided.
- 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.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposes BeerLaNet, an effective stain normalization method based on the algorithmic unrolling of Beer-Lambert law and NMF, but it lacks significant methodological novelty, comprehensive comparisons, and code availability, thus limiting its current impact.
- 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.
I would rate the paper as a weak accept, primarily due to its novel contributions and the effectiveness demonstrated through downstream analyses. However, I remain somewhat concerned about the absence of visualization results and the lack of open-source code is a limitation. A response such as “Code will be made public once the paper is accepted” adds limited value, as transparency and reproducibility are important for proper evaluation.
Review #4
- Please describe the contribution of the paper
The paper proposes BeerLaNet, a trainable stain normalization model based on the unrolling of a physics-informed NMF algorithm guided by the Beer-Lambert law. It learns stain-invariant features without requiring prior knowledge of staining protocols and can be integrated into any deep learning pipeline for tasks like classification or detection. The method is evaluated on public and curated datasets, showing improved cross-domain performance over existing approaches.
- 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 clinically relevant problem in digital pathology, namely the variability in stain appearance across datasets, which affects the generalizability of deep learning models. A key strength lies in the novel formulation of BeerLaNet, which combines a physics-informed model (based on the Beer-Lambert law) with algorithmic unrolling of non-negative matrix factorization (NMF), allowing for trainable and stain-invariant color normalization across a wide range of staining protocols.
The methodological framework is solid and innovative, and the authors provide a comprehensive and well-motivated description of the datasets used, including clear explanations of data selection, labeling processes, and resolution of ambiguities through expert consensus.
Another major strength is the diversity and quality of the datasets, which include multiple publicly available and internally curated datasets that capture a wide range of patient and disease variations, increasing the credibility and generalizability of the results.
The experimental evaluation is thorough, covering two distinct downstream tasks—object detection and classification—which highlights the versatility of the method. The chosen metrics are appropriate, and although no formal statistical significance tests are reported, the consistency and depth of the results support the conclusions.
Finally, the authors provide a critical analysis of existing methods, clearly identifying their limitations and showing how BeerLaNet overcomes them, making a substantial contribution to the field. They also briefly discuss directions for future improvement in the conclusion.
- 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.
I did not identify any significant weaknesses in the paper. On the contrary, the authors dedicate attention to aspects that are often overlooked in conference papers but are essential for journal-level contributions. In particular, the detailed description of the data annotation process and the management of labeling ambiguities through expert consensus are commendable and contribute to the overall robustness and credibility of the work.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My decision of strong accept is based on the relevance of the topic, the high quality and clarity of the writing, the completeness of the experimental evaluation, and the significance of the results.
- 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 thank the reviewers for their feedback and are glad they appreciated our model’s methodological contribution (R2,R4), clinical relevance (R1,R2), and experimental validation (R2,R3,R4).
[R1,R3] Performance Advantage/Choice of Foundation Models: Our method is general and can be applied in front of any backbone, and consistently offers performance advantages over existing methods. We demonstrate improvements in multiple detection and classification tasks. Specifically, we observe consistent performance across tasks with our method, whereas baseline methods exhibit substantial variability. Even in the rare scenarios where our method is not the absolute top performer, the performance difference remains minimal, underscoring robustness and reliability. Models like UNI/CONCH/Phikon are mainly for WSI-level representation learning, whereas our tasks focus on cell classification/detection. While DinoBloom is cell-level and could be explored as a backbone, we intentionally adopt standard backbones to isolate our model’s performance.
[R4] Novelty: To our knowledge, there is no deep-unrolling method for stain normalization problems, particularly in the context of physics-based models. Our approach uniquely combines: (1) physics-informed modeling, (2) adaptive low-rank+sparse factorization without templates, and (3) end-to-end trainable unrolling tailored to histology stain normalization.
[R3] Time Efficiency: Our approximate sparse encoding via a fixed number of unrolled steps is demonstrably faster than many existing methods (e.g., Vahadane solves a full sparse encoding for each data point). The most computationally demanding part is the forward pass through linear operators (computing the gradient steps w.r.t. S and D), which we view as a small cost for the performance gain.
[R1,R3] Lack of Ablation Studies: For general ablation to benchmark components, there is little to be done, as our model is a single layer before a backbone. For the number of factorized components, r, we did explore this but did not report the results. In particular, there was little change in performance once r becomes sufficiently large (e.g., >= 8) due to the low-rank regularization on the factors which adapts the rank of the factorization to the data/task.
[R3] Consistency/Interpretability of Channels: While channels are not explicitly assigned to individual stain components, our decomposition is based on the fundamental structure induced by the physics of the image formation process. The spectra initialization (S matrix) is fixed during inference, ensuring each color channel captures consistent features before being passed to the backbone network.
[R3,R4] Normalization is Task Specific/ No Visualizations of RGB: Our method can map back to visually intuitive RGB images through a post-processing step, but we emphasize the primary focus is domain generalization, not visual realism. Our high-level goal is to provide a stain-invariant learning method that improves domain generalization for downstream tasks, rather than strictly maintaining RGB visual fidelity.
[R3] Updates for Gamma/Lambda: In Alg 1, gamma/lambda are learnable via standard backpropagation from the overall task-specific loss function, combined with non-negative projection after the optimization update to ensure they remain non-negative, as described in the submitted manuscript.
[R3] Existing Methods Could Use Learnable Spectra Matrices: It was not clear to us exactly what was being proposed in this comment. The spectra matrix for synthesizing RGB images is just a matrix multiplication, so one could make this a learnable parameter for these methods, but this just adds a linear layer to the front of the backbone, which we do not anticipate to have a sizable impact as in terms of expressivity this can be easily absorbed by other weights within the backbone.
[R1,R4] Code/Reproducibility: The manuscript states, “Code will be made public once the paper is accepted.”
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The paper proposes BeerLaNet, an adaptive stain normalization method that unrolls Nonnegative Matrix Factorization (NMF). Reviewers provided a diverse range of scores, from 2 to 6, though their comments are not necessarily in conflict. For example, R2 praised the methodological contribution of the paper, while R3 raised critical questions regarding methodological details. In particular, it is necessary to clearly demonstrate the advantages of this network over previous NMF-based stain normalization methods (e.g., Vahadane [18]), not only in terms of quantitative performance but also in qualitative comparisons and processing time. Additionally, R1’s suggestion to extend from ResNet-18 to recent pathology foundation model embeddings is also appropriate. Authors need to address these concerns in their rebuttal.
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper presents an approach that integrates the Beer-Lambert model with downstream tasks via algorithmic unrolling. While this integration is novel in formulation, several concerns raised by the reviewers (I have also read the paper in a lot of detail due to the mixed reviews) were not sufficiently addressed in the rebuttal.
First, as R4 rightly noted, the core idea of algorithmic unrolling involving the Beer-Lambert law and NMF is incremental rather than fundamentally novel. The use of physics-informed models and the exact NMF formulation (Vahadane et al.) has been standard in traditional stain normalization for almost a decade. The key question — to what extent the model goes beyond simple unrolling and achieves meaningful integration into modern deep learning frameworks — remains unanswered. This point is crucial for understanding the real contribution of the work.
Second, and more critically, the evaluation design is insufficient to support the claimed contributions. The downstream baseline methods (especially yolov5 trained on 2,638 single cell-only images from scratch) are significantly outdated and weak by today’s standards. Stronger, modern baselines and datasets such as CellViT++, PanNuke, and MIDOG are now widely used, and their absence makes it difficult to fairly judge the model’s performance in a modern pathology context.
While the authors argue that the method is intended for domain generalization rather than stain normalization, they do not provide sufficient evidence to support either claim. The lack of qualitative visualization (e.g., stain-normalized images) or quantitative benchmarking against foundation models (e.g., CellViT++, which incorporates stain invariance) further undermines the evaluation.
The paper makes broad claims — for instance, that it “extends to arbitrary staining protocols, learning stain-invariant representations of images” — without experimental results to support such generalization.
In conclusion, due to the unclear positioning of the contribution and serious limitations in the experimental validation, I recommend rejection at this time. However, I believe the proposed direction is promising, and if the authors can provide rigorous evaluations and more clearly articulate and prove the novelty beyond incremental extensions, future versions of work could make a meaningful impact in the future.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A