Rethinking Semi-Supervised Medical Image Segmentation: A Variance-Reduction Perspective

For medical image segmentation, contrastive learning is the dominant practice to improve the quality of visual representations by contrasting semantically similar and dissimilar pairs of samples. This is enabled by the observation that without accessing ground truth labels, negative examples with truly dissimilar anatomical features, if sampled, can significantly improve the performance. In reality, however, these samples may come from similar anatomical regions and the models may struggle to distinguish the minority tail-class samples, making the tail classes more prone to misclassification, both of which typically lead to model collapse. In this paper, we propose ARCO, a semi-supervised contrastive learning (CL) framework with stratified group theory for medical image segmentation. In particular, we first propose building ARCO through the concept of variance-reduced estimation and show that certain variance-reduction techniques are particularly beneficial in pixel/voxel-level segmentation tasks with extremely limited labels. Furthermore, we theoretically prove these sampling techniques are universal in variance reduction. Finally, we experimentally validate our approaches on eight benchmarks, i.e., five 2D/3D medical and three semantic segmentation datasets, with different label settings, and our methods consistently outperform state-of-the-art semi-supervised methods. Additionally, we augment the CL frameworks with these sampling techniques and demonstrate significant gains over previous methods. We believe our work is an important step towards semi-supervised medical image segmentation by quantifying the limitation of current self-supervision objectives for accomplishing such challenging safety-critical tasks.1

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer SMeb7/10 · confidence 3/52023-07-01

Summary

Edit: Updating score from 6 t o7 based on the discussions. Extracting useful representations from unlabelled data for label efficient training of medical image segmentation task is a widely studied problem. This work approaches learning useful representations using contrastive learning (CL) within a semi-supervised setting. Strategies for obtaining variance reducing pixel partitions for CL are presented, along with theoretical analysis that show their variance reduction properties. The variance reduction estimation is also used to improve training stability and convergence. Comprehensive experiments on multiple medical imaging and computer vision datasets are performed showing strong performance improvements compared to other SSL methods. The authors show their method is label efficient across all the datasets.

Strengths

* Focusing on variance reduction guarantees to extract contrastive samples for pixel level training is a strong contribution of this work. * The theoretical analyses showing their unbiasedness and use of variance reduction techniques to improve training stability can have important applications in related domains like self-supervised learning. * The experimental evaluation is extensive, with strong performance improvements on multiple datasets. * Contrastive loss landscape visualisation in Fig 3 is insightful; that the contrastive learning holds across datasets is quite convincing.

Weaknesses

* **Robustness**: This work makes two main claims about the usefulness of their CL framework. Firstly, and convincingly so, about label efficiency. There are several places in the paper where model robustness is alluded to, or strong claims made, without any evidence. This could be because the authors view robustness simply to be good performance across multiple datasets? If so, this should be clarified. Currently, the claims about model robustness are misleading. See [1,2,3] for different robustness analyses of deep neural networks. * **Assumption in Th. 3.2**: The guarantees of $Var[H_{SG}] < Var[H_{NS}]$ only holds if different $P_m$ do not have the same expected value over the aggregation function h(x;p). How is this ensured? In medical images, there are scenarios where the differences between classes are small in both intensity and feature spaces? How does the variance reduction guarantees hold in such situations? * **Aggregation functions**: Aren't the aggregation functions some type of a distance measure? And why is it expensive to compute these on dense pixel grids (L213)? * **Main method in Appendix**: While I appreciate all the details presented in this paper, moving the main method to the Appendix is not a good idea. Several of the important details are in the Appendix and does not serve the purpose of what Appendices are supposed to be. * **Literature overview**: I was curious as to why the authors refrained from discussing self-supervised representation learning both when motivating the work, and also in their general discourse. [1] Bastani, Osbert, et al. "Measuring neural net robustness with constraints." Advances in neural information processing systems 29 (2016). [2] Carlini, Nicholas, and David Wagner. "Towards evaluating the robustness of neural networks." 2017 ieee symposium on security and privacy (sp). Ieee, 2017. [3] Singh, Gagandeep, et al. "Fast and effective robustness certification." Advances in neural information processing systems 31 (2018).

Questions

See Weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Authors do not discuss any limitations of their work.

Reviewer LvuQ7/10 · confidence 5/52023-07-01

Summary

The authors present ARCO, a novel semi-supervised contrastive learning framework that employs a stratified group sampling strategy (i.e. **SG** and **SAG**) to compute gradient estimators with reduced variance, thereby enhancing representation learning in dense contrastive learning. By improving dense contrastive learning, ARCO addresses issues related to class imbalancedness and enhances the performance of semi-supervised segmentation, particularly in scenarios with long-tail distributed anatomical classes. The authors provide theoretical evidence demonstrating the effectiveness of the proposed sampling techniques in reducing the variance of the aggregation function, specifically the contrastive loss. The efficacy of the proposed sampling technique is validated on eight 2D/3D benchmark datasets with different label settings, further reinforcing its effectiveness and practical applicability.

Strengths

The motivations are clear, and the method is reasonable. This study shows me a new insight/direction to consider medical image segmentation. The following are my detailed comments. --- **Empirical contribution** Pixel/voxel-wise sampling constitutes a critical facet of contrastive learning at the pixel/voxel level. With the aid of variance-reduction estimation, the authors proffer two pragmatic approaches - Stratified Group (SG) and Stratified-Antithetic Group (SAG), tailored for pixel/voxel-level segmentation tasks with exceedingly scarce labels. * The authors introduce a novel framework termed ARCO (strAtifed gRoup COntrastive learning) devised for multi-class segmentation tasks. This framework appears to be both intriguing and efficacious. The authors undertake a rigorous validation of the proposed methodologies across eight benchmark datasets, encompassing three 2D medical image segmentation, two 3D medical image segmentation, and three semantic segmentation benchmarks. * The empirical results, both quantitative and qualitative, attest to the efficacy of the proposed model across all label ratios and datasets. For instance, the model demonstrates a marked enhancement in segmentation accuracy (up to 4.1% absolute improvements in Dice coefficient) on the challenging multi-class MMWHS dataset under a 1% label setting. * The authors conduct comprehensive ablation studies to substantiate that the proposed mechanisms merit consideration. These studies encompass eight benchmark datasets, diverse network architectures, and varying label ratios to validate the efficacy, model-agnostic nature, and label efficiency of the proposed methodology. * Lastly, the proposed methodologies are not only facile to implement but also boast of universal applicability. For instance, they can be seamlessly integrated into any scenario necessitating pixel/voxel sampling. The paper presents a robust and versatile framework for pixel/voxel-level contrastive learning, which is empirically validated through extensive experiments and ablation studies. The methodologies are characterized by ease of implementation and broad applicability, making them a valuable contribution to the field of image segmentation. --- **Theoretical contribution** * The proposed methodologies, SG (Stratified Group) and SAG (Stratified-Antithetic Group), have exhibited remarkable efficacy in the experimental study. Consequently, it is intriguing to ascertain whether theoretical insights can elucidate this enhancement in performance. To this end, the paper furnishes a cogent theoretical analysis of the methodologies, revealing that the variance-reduction attribute of the two sampling methods is instrumental to their performance. * First and foremost, in Section 3.3, the paper meticulously delineates the SG and SAG sampling methodologies through lucid mathematical equations. SG is executed by segregating pixels into mutually exclusive groups, followed by uniform sampling of a specified number of pixels from each group. SAG, which is predicated on SG, imposes an additional constraint of symmetry among the sampled pixels within each group. * Subsequently, the sampled pixels are amalgamated through an aggregation function, which acts as an estimator for the target quantity. In the realm of image segmentation, this quantity could be, for instance, the contrastive loss function. It is posited that an optimal balance must be struck in the sample size; an overly diminutive sample size may fail to encapsulate the salient information from the underlying image, whereas an excessively large sample size would entail high computational complexity. Thus, the ideal scenario would be for SG to capture the crux of the image information through a relatively modest sample size. The paper demonstrates that SG possesses this attribute by establishing that it achieves reduced variance in comparison to the naïve sampling method (i.e., uniform random sampling from all pixels). Specifically, Theorem 3.2 establishes that SG is an unbiased sampling methodology, with a variance that does not exceed that of naïve sampling. More precisely, the variance of SG can be decomposed into the variance of naïve sampling minus a non-negative term. This non-negative term is conjectured to be almost certainly greater than zero, as it would be zero only if the expectation of the aggregation function within each group is identical to the expectation over the entire image, which is highly improbable. Thus, Theorem 3.2 suggests that SG is likely to consistently outperform naïve sampling. * Figure 5 reveals that SG/SAG exhibits marginally expedited training convergence compared to naïve sampling. This observation is theoretically substantiated in the concluding paragraph of Section 3, which is commendable. * In my assessment, the theoretical analysis presented in the paper is inextricably linked to the empirical component and provides a persuasive rationale for the empirical performance augmentation of SG vis-à-vis naïve sampling, thereby bringing the empirical narrative full circle. --- **To sum up** **1. Clarity** The manuscript is eloquently composed, proffering a lucid and cogent progression of information. The authors adeptly elucidate the procedural framework, facilitating the readers' comprehension of the proposed methodologies, namely, the two instance sampling methods - Stratified Group (SG) and Stratified-Antithetic Group (SAG). In Section 3.3, the SG and SAG sampling methods are meticulously delineated through precise mathematical formulations. The discourse within the Methodology section (Section 3) furnishes an exhaustive exposition of the innovations introduced by the study, adeptly accentuating the distinct contributions of the research to the scholarly domain. **2. Novelty** This work is the first work, empirically and theoritically, to validate the variance-reduction approach within the context of pixel/voxel-level contrastive learning for semi-supervised medical image segmentation, particularly in scenarios characterized by a paucity of labels. **3. Experimental Comprehensiveness** The authors undertake a comprehensive suite of experiments encompassing eight medical datasets, which include both 2D and 3D modalities, as well as semantic segmentation benchmarks. Additionally, a diverse array of contrastive learning frameworks and varying label ratios are employed to rigorously assess the efficacy, model-agnostic properties, and label efficiency of the proposed methodology. This extensive experimental evaluation substantiates the robustness and versatility of the technique in the domain of medical image segmentation. **4. Theoretical Implication** The authors furnish a cogent and meticulously articulated theoretical analysis of the proposed approach, elucidating the underlying principles with clarity and precision. This analytical exposition contributes to a deeper understanding of the methodology's foundations and its implications.

Weaknesses

The proposed method is very interesting. There is no obvious weakness in the proposed ARCO. However, I do have a few questions. See the following section.

Questions

* The proposed two sampling techniques are intriguing. I wonder if you could kindly provide insights on the applicability of these sampling techniques to different tasks. Such information would prove beneficial not only to the current domain but also to other related domains and scenarios. Understanding the potential applications of these techniques would undoubtedly contribute to the advancement of various fields. * Lemma 3.1 demonstrates that the variance of SAG is at most twice of that of SG, which implies that the variance reduction magnitude of SAG and SG are roughly at the same level. Could the authors further elaborate on when SAG be preferred to SG?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors have addressed the potential broader impact in clinical scenarios.

Reviewer QWqe5/10 · confidence 3/52023-07-02

Summary

This paper proposes two new sampling strategies, SG and SAG, for contrastive learning in semi-supervised frameworks. Compared with randomly sampling pixels for contrastive learning, pixels are grouped into several subsets, and then pixels are sampled from each subset. The proposed method is proven to reduce the variance of sampled pixels. Solid experiments are conducted to support the above claim.

Strengths

1. Solid experiments and good performance. 2. Simple yet effective method that benefits contrastive learning for semi-supervised frameworks. 2. Sound theoretical analysis.

Weaknesses

1. It would be better to also compare NS with the proposed SG/SAG in natural image datasets. 2. What is the application scenario of SAG? SG seems to have better performance and stability than SAG. 3. Line 222-223, what is the meaning of "p is orthogonal to p'"?

Questions

See weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have included a discussion on the potential negative societal impact.

Reviewer J4c54/10 · confidence 4/52023-07-04

Summary

The authors propose a sampling strategy to improve contrastive-learning-based medical image segmentation performance with limited labeled data and training stability. The sampling strategy is used in conjunction with a previously-published contrastive semi-supervised training strategy. The main contributions include this sampling strategy, some theoretical support that the strategy should improve training stability, and experiments evaluating the method.

Strengths

- The baseline comparison experiments are thorough. - The authors address a real applied problem (that most medical image segmentation datasets have limited data) with a new method and theoretical support, resulting in a well-rounded paper.

Weaknesses

Major weaknesses - The paper’s writing is a major limitation. The prose is difficult to understand, which limits the entire paper—it is hard to clearly understand the motivation, the proposed method, the contributions, or the benefits. The work would benefit from more rounds of grammatical revision; it is hard to parse what the author is trying to communicate a lot of the time. - Partly due to the writing issues, it is unclear how the proposed sampling strategy differs from previous dense contrastive learning-based image segmentation strategies; it seems that the methodological contributions here are minor, if existent. - The method is complex, consisting of two backbones (connected via EMA), global/local instance discrimination losses, augmentations, supervised losses, nearest neighbor loss, global contrastive loss, unsupervised loss… with all of these components, a very thorough ablation section is needed to understand how much the novel component (a pixel-level sampling strategy) is contributing to performance. The existing ablation section is not so thorough. As a result, the paper does not contribute much understanding about the strengths/weaknesses of different components of this complex pipeline. Minor weaknesses - MONA is not a well-known training strategy; it would be useful to provide a longer overview on what MONA does and how the proposed approach differs. This discussion could go in an appendix. - The same concept is often referred to using different words: model “convergence,” “robustness,” and “stability.” I’m not sure if you’re always talking about the same concept, or if you are using different words to refer to the same idea. If the latter, it helps the reader to always use the same word. - This is a minor stylistic note, but the use of bold and italics is often distracting.

Questions

Questions and suggestions discussed above

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

1 poor

Contribution

2 fair

Limitations

Limitations and impacts adequately addressed.

Reviewer J4c52023-08-11

Rebuttal response

I have read the other reviews and rebuttal responses and I appreciate the authors’ thorough rebuttal. I have reread the paper as well during this rebuttal period. To summarize my thoughts: - I think the paper shows promising empirical results at training segmentation models with few labeled images. I am encouraged by the authors’ addition of of new ablation experiments. - I think reviewer SMeb brings up a good point about the term and associated claims of “robustness” that I did not include in my original review. - I still find the paper very difficult to interpret and the proposed training method to contain so many components that it is difficult to know which components are useful for other pipelines. I recognize that the other reviewers do not seem to have the same perspective. I will change my score from a 3 to a 4 due to the additional experimental support during this rebuttal period, but I would still vote to reject this submission. I would point the AC to Sections 3.2 and 3.3 of the paper, which contain the paper’s contributions, if they wish to investigate further and see if they find the same problems I do with the submission or if they are more aligned with the other reviewers.

Authorsrebuttal2023-08-13

Response to Reviewer J4c5 (Part 1)

Thank you for taking the time to re-examine our paper during the rebuttal phase! We'd like to further address your additional queries as follows: > **Q1**: I think the paper shows promising empirical results at training segmentation models with few labeled images. I am encouraged by the authors’ addition of of new ablation experiments. **A1**: We appreciate your recognition of the paper's promising empirical results and our commitment to improving the manuscript with the addition of new ablation experiments. If you have further concerns on ablation studies, please feel free to contact us. > **Q2**: I think reviewer SMeb brings up a good point about the term and associated claims of “robustness” that I did not include in my original review. **A2**: Many thanks! We appreciate the opportunity to provide clarity on the term "robustness": In this context, “robustness” is a comprehensive concept used to describe the performance of a segmentation model. A model is said to be robust if (1) it has a high segmentation quality with only using extremely limited labels in long-tailed medical data (please see Line 141); (2) and fast convergence speed (please see Line 207). Our introduced sampling methods, namely Stratified Group (SG) and Stratified-Antithetic Group (SAG), have shown improvements in segmentation performance across different pipelines. Furthermore, these methods can seamlessly integrate with existing frameworks, further enhancing the robustness (as detailed in Appendix Lines 825-840, Section L and Appendix Page 25, Table 6). We greatly appreciate the feedback and will make it a priority to refine and highlight this to improve the paper's clarity.

Authorsrebuttal2023-08-13

Response to Reviewer J4c5 (Part 2)

Thanks for your helpful comments! If you have further concerns, please feel free to contact us. > **Q3**: I still find the paper very difficult to interpret and the proposed training method contains so many components that it is difficult to know which components are useful for other pipelines. I would point the AC to Sections 3.2 and 3.3 of the paper, which contain the paper’s contributions. **A3**: We sincerely value your feedback and will thoroughly revise the section in our final revision. We first clarify that we do **NOT** claim the training pipelines as our contribution. Instead, we have introduced two plug-and-play sampling techniques that seamlessly integrate with existing state-of-the-art Contrastive Learning (CL) frameworks adopting the student-teacher design. Section 3.2 discusses the potential high variance in CL frameworks and presents strategies for addressing these challenges along with their implementation details. Section 3.3 offers a theoretical analysis explaining the efficacy of our proposed method in enhancing variance reduction, accompanied by optimization guarantees (detailed further in Appendix, pages 595-647, Sec. A). Below, we would like to offer further clarification on our motivation and contributions. #### **Major Motivations**: Intuitively, using Contrastive Learning (CL) aids in learning robust representations for medical segmentation if positive and negative pairs align with the intended latent anatomical classes [46,13]. However, recent studies suggest potential issues, such as representation collapse into constant features [11,10,16] or confinement to a lower-dimensional subspace [17,18,19,20]. A key contributor to this fragility might be the non-smooth feature space near samples [21,22], where sampling of pixels could introduce significant feature variations. Existing sampling methods, such as naive sampling, often produce higher variances and struggle to discern semantically similar pixels [21], potentially compromising CL's stability. This prompts the pivotal query: How might we adeptly select the most informative pixels/voxels, thereby reducing variance in semi-supervised CL models? As illustrated in Figure 3 (Lines 217-218), regions with similar anatomical features within medical images should cluster, yielding corresponding plateaus in the loss landscape visualization, a perspective echoed in recent empirical studies [66,67]. #### **Key Contributions**: - **Technical Novelties**: We introduce two sampling techniques, i.e., Stratified Group (SG) and Stratified-Antithetic Group (SAG), to mitigate the undesirable high-variance limitation through the concept of variance-reduced estimation. - **Robustness Across Different Pipelines**: Through extensive experimental results, we demonstrate that two proposed sampling methods combined with state-of-the-art CL frameworks perform well in pixel/voxel-level segmentation tasks with extremely limited labels (further detailed in Appendix Lines 825-840, Section L, and Appendix Page 25, Table 6). - **Theoretical Analysis**: To our best knowledge, we are the **first work** to show the benefit of certain variance-reduction techniques in CL for medical image segmentation. We unveil the untapped potential of the refined gradient estimator for handling long-tailed medical image data. - **Comprehensive Experiments**: We conducted extensive experiments on **eight** diverse 2D/3D medical and semantic segmentation benchmarks, demonstrating the efficacy of our approach across various label ratios, i.e., 1\%, 5\%, 10\% (further detailed in Lines 261-319 Section 4.1, Line 247-248 Table 1, Line 273-274 Figure 1, and Appendix Page 22-35, Sections G-K). Thank you for your continued consideration and time! Please feel free to contact us for further concerns.

Authorsrebuttal2023-08-21

Look forward to further feedback

Dear Reviewer J4c5: As the review period draws to a close, we would like to emphasize our gratitude for your insightful comments, which have significantly shaped our revisions. We wish we have successfully addressed your concerns. We sincerely value your feedback and insights and humbly request your endorsement. We wish to emphasize our meaningful contribution to the realm of semi-supervised medical image segmentation. Our work also provides the community with significant insights into generic vision intelligence, especially in understanding the nuances of current self-supervision objectives for complex, safety-critical tasks. We believe our research can substantially enrich the collective knowledge in this domain. With a brief window remaining for discussion, we remain open to further dialogue to address any residual questions or concerns. Your continued engagement is highly regarded and vital to our research aims. Thank you once again for your dedication and thoughtful consideration. Best Authors of Paper1499

Reviewer LvuQ2023-08-14

Response to Authors' rebuttal

Thank you for your responses to the questions I raised. The response was satisfactory and addressed my concerns. Besides that, I'd also like to discuss/highlight the following point. - The notable contribution of the proposed SG sampling method in improving the representation quality for pixel/voxel-level contrastive learning. The superiority of SG over na&iuml;ve sampling has been shown across various datasets in the draft. Nonetheless, I would appreciate further insight into the SAG method. From my observation, SAG does not surpass SG in most experiments, except in the specific context of the SUN RGB-D benchmark under 50-label setting. Yet, the performance of SAG remains comparable to SG. Consequently, I interpret the SAG method as an alternative to SG, designed to achieve comparable results but with a reduced sample size. Would this be a correct interpretation?

Authorsrebuttal2023-08-14

Response to Reviewer LvuQ

We sincerely thank the reviewer for acknowledging the positive changes we have made to the paper. If you have further concerns, please feel free to contact us. > **Q1**: The notable contribution of the proposed SG sampling method in improving the representation quality for pixel/voxel-level contrastive learning. The superiority of SG over naïve sampling has been shown across various datasets in the draft. Nonetheless, I would appreciate further insight into the SAG method. From my observation, SAG does not surpass SG in most experiments, except in the specific context of the SUN RGB-D benchmark under 50-label setting. Yet, the performance of SAG remains comparable to SG. Consequently, I interpret the SAG method as an alternative to SG, designed to achieve comparable results but with a reduced sample size. Would this be a correct interpretation? **A1**: We thank the reviewer for acknowledging the positive changes we have made to the paper. Yes, your interpretation is correct! SAG can halve the sample sizes compared to SG, while largely preserving SG's variance reduction property, offering enhanced theoretical efficiency. We thank the reviewer again for the constructive feedback which helps shape this revision! Please do not hesitate to reach out should you have any additional feedback or questions.

Authorsrebuttal2023-08-16

Thank you again for your review and very valuable feedback!

We deeply appreciate the reviewer's recognition of the improvements made to our paper. Once again, we extend our gratitude for the invaluable feedback that has greatly contributed to refining our work!

Reviewer SMeb2023-08-16

Response to author rebuttal

I have now read the author rebuttal, which addresses most of the concerns raised in my initial review. I have also seen the discussions between other reviewers and the effort invested by authors in these discussions is appreciated. I am willing raise my score from 6 t o7.

Authorsrebuttal2023-08-16

Thank you again for your review and very valuable feedback!

We sincerely thank you for taking the time to review our rebuttal and for recognizing the positive changes we have made to address the concerns raised. We deeply appreciate your consideration in adjusting the score. Your feedback has been invaluable in refining our work, and we are committed to ensuring the highest quality in our final manuscript. Again, we express our gratitude for your thoughtful review and positive reassessment!

Area Chair Tp4X2023-08-18

Rebuttal to QWqe

Dear QWqe Could you have a look at the rebuttal to see if your questions have been clarified? Thanks, Your AC

Authorsrebuttal2023-08-20

Look forward to further feedback

Dear Reviewer QWqe: We are genuinely thankful for your thoughtful feedback, which has been pivotal in refining our manuscript. As the author-reviewer discussion period is nearing its conclusion, we kindly request your review of our rebuttal and any further reflections you might have. Please feel free to indicate any additional clarifications or experiments that could further strengthen our paper. We aim to unequivocally convey the significance of our work. If you feel that our responses have adequately addressed your concerns, we would be grateful if you might consider raising the paper's rating. Once again, we deeply value your comprehensive review and are thankful for the positive evaluation. Your feedback has been pivotal in refining our work, and we are committed to incorporating all of the suggestions into our manuscript. Best Authors of Paper1499

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC