In image recovery problems, one seeks to infer an image from distorted, incomplete, and/or noise-corrupted measurements. Such problems arise in magnetic resonance imaging (MRI), computed tomography, deblurring, super-resolution, inpainting, phase retrieval, image-to-image translation, and other applications. Given a training set of signal/measurement pairs, we seek to do more than just produce one good image estimate. Rather, we aim to rapidly and accurately sample from the posterior distribution. To do this, we propose a regularized conditional Wasserstein GAN that generates dozens of high-quality posterior samples per second. Our regularization comprises an ℓ1 penalty and an adaptively weighted standard-deviation reward. Using quantitative evaluation metrics like conditional Fréchet inception distance, we demonstrate that our method produces state-of-the-art posterior samples in both multicoil MRI and large-scale inpainting applications. The code for our model can be found here: https://github.com/matt-bendel/rcGAN.
Paper
Similar papers
Peer review
Summary
This paper introduces a regularization term, comprising an L1 penalty and a standard-deviation reward, for the conditional GAN with Wasserstein loss. The objective is to ensure the generation of high-quality samples for a specific input observation y. The proposed approach is evaluated on two image recovery tasks: MRI reconstruction and face image inpainting. The numerical results illustrate its effectiveness in comparison to alternative methods.
Strengths
1) The proposed regularization term is simple and easy to implement and be added to GAN models; 2) Discussed different regularization options and demonstrated why they did not work in this case; 3) It includes theoretical analysis and empirical results; 4) The overall writing and presentation are good and easy to follow.
Weaknesses
1) The novelty of this paper sounds a little bit limited. The supervision-L1 to regularize conditional GAN is proposed by previous work. So it looks like it only added a new reward term on top of previous work. 2) The two experimental tasks are all linear image restoration tasks. However, it is curious how the proposed method would work on nonlinear tasks. 3) It would also be interesting to see whether the proposed method would work on a more complicated real-world dataset for the inpainting task. 4) The current experimental image datasets are low resolution. Will it still outperform other models in terms of quality and computational cost on higher-resolution image datasets? 5) How to determine the hyper-parameter: belta_adv for different datasets? 6) Minor issue: there are some writing inconsistencies (e.g., the capitalization of the headline of 4.2 /4.3 is different from the headline of 3.1/3.2). #------------------------------------------------------------------------------------------------# #------------------------------------------------------------------------------------------------# Most of my previous concerns have been addressed by the authors during the rebuttal, thus I have increased my rating accordingly. However, it would be great if the authors could add some experiments about non-linear image inverse problems.
Questions
See [Weaknesses].
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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
3 good
Contribution
3 good
Limitations
Yes, the authors have discussed their limitations in the Conclusion section.
Summary
This work uses conditional GANs for solving image recovery problems including face-inpainting and accelerated MRI. In particular, a regularization scheme is proposed that tackles the lack-of-diversity issue related to posterior sampling from cGANs. The authors show that the proposed regularization allows sampling from the true posterior under the assumption of a Gaussian signal prior. Empirically the method outperforms a U-Net trained end-to-end, existing cGANs and two score/diffusion/Langevin approaches in terms of accuracy and the latter also significantly in terms of speed.
Strengths
1. The paper is written very clearly and is well organized. 2. The proposed regularization significantly improves over existing cGAN approaches and hence will be of interest for the community. 3. The theoretical discussion of a toy example with Gaussian signal prior is insightful and contributes to the understanding of the proposed regularization.
Weaknesses
1. The authors argue that sampling from the posterior distribution alleviates any concerns about fairness and facilitates uncertainty quantification and further that uncertainty quantification in MRI, and fairness in face-generation, are both of paramount importance. In Appendix A they show that a classifier trained on images from the prior distribution maintains its performance when averaged over many samples from the posterior but not when evaluated on a point estimate. However, the empirical experiments in Section 4 do not evaluate these points in detail. Regarding uncertainty quantification for MRI Figure 4 shows one example of the pixel-wise standard deviation (SD) but a discussion on the usefulness and practical applicability is missing. To me it seems that in the depicted SD map the model is equally uncertain about every edge in the image making me wonder about the usefulness of this map to a radiologist. Regarding fairness in face-generation 5 different reconstructions of one image are shown and the authors argue that the fact that some samples show almond-shaped eyes demonstrates fairness. Here, a more detailed evaluation would strengthen the contribution although I acknowledge that this might be out the scope for a single paper. 2. In terms of reconstruction accuracy the U-Net is shown to be outperformed but I would assume that a stronger baseline trained end-to-end like a Var-Net https://arxiv.org/abs/1704.00447 would perform significantly better in terms of PSNR and SSIM. Further, the conditioning makes the model lose the advantage of unconditioned generative models for solving inverse problems to be independent from the forward map. Since the authors also do not convincingly demonstrate cGANs to be superior in terms of fairness and uncertainty estimation (see previous pont), I wonder about the general usefulness of cGANs as proposed in this work. However, I acknowledge that this work significantly outperforms previous cGAN based approaches and thus is a valuable contribution to the field of generative models for solving inverse problems. Also the authors discuss the dependence on a specific forward model as a limitation and refer to ongoing work that removes this constraint.
Questions
No additional questions.
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
4 excellent
Contribution
3 good
Limitations
Limitations are discussed adequately by the authors.
Summary
The paper presents a new method based on conditional GANs (cGANs) for generating posterior samples for solutions of inverse problems. The motivation comes from the fact that standard approaches to inverse problems typically use point estimates, which due to the perception-distortion tradeoff will lead to blurry images. Conversely, reconstruction via sampling can be used to navigate the perception-distortion tradeoff by either sampling small counts to have sharp images, or averaging many samples to approximate the point estimate reconstruction. All of the above is known from the previous literature. The present paper's specific contributions largely center on the loss function and regularization design for cGANs to avoid posterior collapse and increase sample quality. The regularizer design is based on matching the means and covariances of the samples to the means and covariances of the true distribution. The manuscript chooses to compute these statistics over the $\ell_1$ norm rather than the $\ell_2$ norm, and the paper shows that $\ell_2$ norm can in fact lead to collapse. In numerical experiments the paper demonstrates superior FID and conditional FID performance for the proposed method, indicating that it is superior at distribution matching than the previous methods considered. For distortion metrics, the experiments demonstrate that superior distortion performance can be achieved by averaging over samples.
Strengths
**Originality** - Interesting new regularization method for training cGANs for image restoration tasks. **Quality** - The paper's quantitative results indicate SotA performance across a variety of tasks for both MRI reconstruction and image inpainting. - Principled use of VGG weights instead of Inception weights for FID calculation and correlations to radiologist perception. - A derivation motivating the use of both the supervised standard deviation reward and the application of this award over $\ell_1$ norms. **Clarity** - Excellent presentation of perception-distortion tradeoff and its relevance to imaging inverse problems. - Detailed outline of the proposed method with the appendices, as well as relevant mathematics concerning the various tasks in the results. **Significance** - There is substantial interest in the image reconstruction/restoration community now, much of which has focused on conditional diffusion models. The establishment of competitive performance in this space with cGANs (as well as the superior performance to past cGAN methods) will be of interest.
Weaknesses
**Quality** - The MRI images still demonstrate substantial artifacts from all methods at 8X acceleration. It is possible to find a number of features and fine details that are missing, so I doubt these would be accepted for clinical use. - Although VGG may correlate better with radiologist scores, it is still an ImageNet feature backbone and contains nullspaces that may be poor for the tasks under consideration (Kynkäänniemi, 2023). - There is a literature dedicated to so-called "plug-and-play" methods that are relevant to the present work, but was not considered. See (Venkatakrishnan, 2013) and follow-up papers. - The U-Net is used as a surrogate for point-estimate methods in the MRI example, but it is known in the MRI community that the U-Net is a poor reconstructor. A more standard comparison would be to use the End-to-End Variational Network (Sriram 2020), which is available from the fastMRI GitHub repository. This comparison was included in the cited paper of (Jalal 2021). - The competing methods for inpainting seem out of date. Dozens of methods based on conditional probabilistic diffusion have emerged over the last year (e.g., DDRM (Kawar 2022) and follow-ups), so I do not find these experiments as strong as the MRI ones. 1. Kynkäänniemi, Tuomas, et al. "The Role of ImageNet Classes in Frechet Inception Distance." ICLR (2023). 2. Venkatakrishnan, Singanallur V., Charles A. Bouman, and Brendt Wohlberg. "Plug-and-play priors for model based reconstruction." 2013 IEEE global conference on signal and information processing. IEEE, 2013. 3. Sriram, Anuroop, et al. "End-to-end variational networks for accelerated MRI reconstruction." Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part II 23. Springer International Publishing, 2020. 4. Kawar, Bahjat, et al. "Denoising diffusion restoration models." Advances in Neural Information Processing Systems 35 (2022): 23593-23606.
Questions
I am positive on this work, but a number of aspects on the presentation could be improved. I would appreciate the authors' responses on improving the following, mostly related to the limitations of the work and contextualization with existing literature on medical image reconstruction. 1. Although sampling-based reconstruction can do a better job of distribution-matching and uncertainty estimates, the artifacts introduced are unpredictable. For medical imaging, calibrated radiologists can work around predictable artifacts such as blurring. In fact, even point-estimate reconstruction has not fully taken over the medical field, because simple linear reconstructions have more predictable artifacts than deep learning point estimates. The use of generative sampling compounds this effect and is likely to be substantial limitation on the present work. I think a discussion on this should be added as limitation (some sketch of the proposed discussion would be appreciated during the rebuttal). 2. The paper primarily examines distribution matching over ImageNet features, which are likely to be quite different than the features useful for medical imaging. (Kynkäänniemi, 2023) shows the presence of perceptual nullspaces that could very well ignore small pathologies. Even the self-supervised solution of (Kynkäänniemi, 2023) is likely to be of benefit for medical imaging analysis, because self-supervised features rely on large objects at the center of the image. 3. Although the classifier matching proof in Appendix A seems theoretically sound, it relies on a few limits and calibration assumptions that are likely to deviate from reality, so I am skeptical this is a grounded proof that classifiers trained on fully-sampled images will perform adequately on the sampled images. 4. The paper could benefit with a comparison to the End-to-End variational network, which is a standard comparison method in the MRI literature and was previously used in the Jalal paper. Although the present manuscript has shown superiority to (Jalal, 2021) on metrics, sampling-based methods have not yet taken the general field of medical image reconstruction, so the comparison is still valid. 5. A cursory overview of the Plug-and-Play reconstruction literature would provide a more complete picture of previous work.
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.
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
4 excellent
Presentation
2 fair
Contribution
3 good
Limitations
As discussed under Questions, the paper could expand its discussion of limitations, including the issues of 1) issues artifact consistency posterior sampling methods with human observers and 2) the use of metrics based on ImageNet feature alignment for medical imaging tasks.
Summary
This paper proposed a cGAN (conditional generative adversarial network) method for general image inverse problems (IR). Conventional cGAN methods for IR usually suffer from mode collapse that the generators learned to ignore latent inputs and are unable to generate diverse image reconstruction samples. Motivated by existing work, this work proposes to incorporate multiple latent input to the GAN model during training to mitigate the mode collapse effect. Accordingly, a sophisticated designed supervised training loss function is also presented in this work. The numerical results show that the proposed method generates comparable or even better results to cGAN and score-based baseline methods on compressive MRI and image inpainting.
Strengths
1, The paper is overall well motivated, clearly written and easy to follow. 2, The proposed posterior sampling-based method is effective for generating multiple samples compared to some score-based methods. The multiple sampling can also be applied for model uncertainty estimation, which is important for many medical imaging problems. 3, The theoretical interpretation on the mean and covariance of predicted posterior distribution is informative, though it is sometimes unrealistic for most real-world imaging problems.
Weaknesses
1, One contribution of this paper seems still to be based on the idea of Adler & Oktem [9], where multiple latent vectors are as inputs to the discriminator and generator module instead of only one input. In the meantime, as pointed out by the author as well, adding extra regularization term in the adversarial GAN loss during training is not completely new. 2, This proposal ignores a body of work for posterior sampling based on Bayesian deep learning such as Monte Carlo dropout, deep ensembles, etc. There’s also a lack of numerical comparisons with these methods, making the baseline methods not sufficient enough. 3, While some ablation studies provided in its original submission, the effects of regularization parameter $\beta_{SD}$ and number $P$ on the final imaging reconstruction are still not comprehensive and could be improved. For example, a visual illustration or numerical comparison of several proposed sampling results due to different settings of $\beta_{SD}$ is preferable. 4, The assumption for proposition 3.1 that samples $\widehat x_i$ and $x$ are independent Gaussian conditioned on y is not true for imaging applications, weekend the theoretical contribution. 5, Some technical details and notations are confusing and could be improved in their current state.
Questions
1), There seems to have a gap between proposition 3.3 and the gradient descent on the $\beta_{SD}$ and could be clarified for reader’s better understanding. For example, there is no clear formular between $\beta_{SD}$ and the ratio in Eq 20., making the derivative of Eq. 21-23 a bit confusing. It is unclear how Eq. 21-22 is formulated and how this led to the update of $\beta_{SD}$. Please clarify this part in the revision. 2), At the same line above, it is unclear why the number of latent $P$ is not the same for training and testing (e.g., line 166,” for some $P_{\text{val}} $ that does not need to match $P_{\text{train}} $.”). Is this means the performance of this proposal for each image is sensitive to the choice of value $P$? It is unclear when there is a lack of enough validation dataset to finetune $\beta$ or if there is a mismatch between train and testing dataset. As a result, the effects of hyper-parameter settings for different settings are unclear. 3), The paper lacks sufficient discussion on the CS-MRI experiments. For instance, based on the description of Sec. 4.2, “we transformed $x_t$ to the Fourier domain”, seems the authors do not follow the trends of other methods in the fast MRI leaderboard that directly perform down-sampling on the kspace data. This seems like an important distinction and must be made clear in the paper. There is no discussion on how the measurement noise is modeled or how it affects the reconstruction performance. Also, while there are some technical discussions on the CSMRI formulation in the appendix, there is no description of how the forward model simulation in this paper differs from the actual forward model of the real MRI system. 4), One benefit of posterior sampling is for model uncertainty estimation. While this work can generate multiple samples and show the pixel-wise standard deviation in Figure 4, it is difficult to evaluate if it can indeed reflect the mean-absolute error for image reconstruction. It would be great to have a quantitative result for this study. 5), There is a lack of comparison to more recent diffusion model-based IR methods, where many have been demonstrated good posterior sampling performance for various imaging tasks, such as [R1, R2, R3]. 6), The run time comparison to the selected Langevin dynamic baselines seems to be suboptimal, as there are many efficient sampling strategies can be applied for IR diffusion/score-based methods, where fewer iteration steps could result in appealing reconstruction [R3]. 7), It is unclear which forward model/dataset are used for Figure 2 in the paper. If it is plotted using images, it is unclear how is the curve in theory plotted. Moreover, it would be more informative if Figure2 can be directly represented by $\epsilon_1/\epsilon_P$ and $\beta_{SD}$. [R1] Score-based diffusion models for accelerated MRI, Chung. etal, 2021. [R2] Diffusion posterior sampling for general noisy inverse problems, Chung. etal, 2022. [R3] Pseudoinverse-Guided Diffusion Models for Inverse Problems, Song. etal, 2023.
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
2 fair
Presentation
3 good
Contribution
3 good
Limitations
The authors discussed their work's limitations in the submission.
Thank you for your response. I have also read the other reviews and responses. However, I still think there is an outstanding issue with the limitations. Although Sriram '20 may have had issues at 8X acceleration, we also know that 8X acceleration is not clinically acceptable for the E2E VarNet, even though it is acceptable for 4X (Recht et al., 2020, Muckley et al., 2021). My point was that the present work still does not seem to resolve the 8X usability issue. The response on the questions of radiologist reading through blurring does not directly address my concern. In general, radiology is extremely risk-averse. It is much easier to just scan a little bit longer than to use to a risky method, and medical practitioners will generally opt for the less-risky method. Uncertainty maps can help, but it is easier to just not deal with the problem. As for the ImageNet feature alignment, this reference, like others in the field, primarily relies on radiologist preference. It can be difficult to evaluate observer performance. For clinical deployment, the critical component is not what images radiologists like, but whether they can see the pathology. Radiologist rankings of aspects such as SNR or CNR often don't correlate to whether they can see pathology - for example, abdominal imaging in general is very bad for SNR, but still shows many pathologies. So, in summary, I still think the limitations should be expanded for both points I made above, to be more conservative in the claims that the improved results probably need to be validated with a pathology-centric radiologist study, and with the current experiments don't assess that performance. If the limitations discussion were expanded to incorporate both of these, I could improve my score. 1. Recht, Michael P., et al. "Using deep learning to accelerate knee MRI at 3 T: results of an interchangeability study." AJR. American journal of roentgenology 215.6 (2020): 1421. 2. Muckley, Matthew J., et al. "Results of the 2020 fastMRI challenge for machine learning MR image reconstruction." IEEE transactions on medical imaging 40.9 (2021): 2306-2317.
We thank the reviewer for the continued discussion and opportunity to improve our submission. - **Reviewer**: Although Sriram '20 may have had issues at 8X acceleration, we also know that 8X acceleration is not clinically acceptable for the E2E VarNet, even though it is acceptable for 4X (Recht et al., 2020, Muckley et al., 2021). My point was that the present work still does not seem to resolve the 8X usability issue. **Response**: We chose to focus on 8X because it yields more uncertainty and thus more interesting images than 4X. But we agree that 8X acceleration may be outside of the clinical acceptability window for our method and perhaps all existing methods. Experimentally, we have observed that our performance gains over our competitors are similar at 4X compared to 8X. However, our submission included only a 4X version of Table 1 (that includes CFID, FID, APSD, and run-time) in Appendix J.1. Thus, if our paper is accepted, we will do the following. *First, we will add a limitation statement acknowledging that at 8X acceleration, the recoveries generated by our method likely fall below clinical quality. Second, we will add a complete reporting of 4X results to our paper's supplementary materials. In particular, we will add a 4X version of Table 2 (that includes PSNR, SSIM, LPIPS, and DISTS) as well as 4X versions of Figure 1, Figure 2, and Figure K.1 through K.6.* - **Reviewer**: The response on the questions of radiologist reading through blurring does not directly address my concern. In general, radiology is extremely risk-averse. It is much easier to just scan a little bit longer than to use to a risky method, and medical practitioners will generally opt for the less-risky method. Uncertainty maps can help, but it is easier to just not deal with the problem. **Response**: Although it's possible to lower the acceleration rate in certain applications (e.g., 2D brain and MSK imaging), it is not possible in others, such as breath-held cardiac cine or dynamic applications in general. For this reason, we believe that it is important to quantify the fundamental uncertainties associated with MR image recovery, and this is the goal of posterior sampling. In any case, we agree that radiology is extremely risk-averse, and so *we will add a limitation statement acknowledging that more rigorous studies are needed before radiologists feel comfortable adopting generative deep-learning-based methods in clinical practice.* - **Reviewer**: As for the ImageNet feature alignment, this reference, like others in the field, primarily relies on radiologist preference. It can be difficult to evaluate observer performance. For clinical deployment, the critical component is not what images radiologists like, but whether they can see the pathology. Radiologist rankings of aspects such as SNR or CNR often don't correlate to whether they can see pathology - for example, abdominal imaging in general is very bad for SNR, but still shows many pathologies. So, in summary, I still think the limitations should be expanded for both points I made above, to be more conservative in the claims that the improved results probably need to be validated with a pathology-centric radiologist study, and with the current experiments don't assess that performance. If the limitations discussion were expanded to incorporate both of these, I could improve my score. **Response**: We agree that the critical component of image quality is whether radiologists can see the pathology. Thus, *we will add a limitation statement acknowledging that our results, as well as the metrics on which we measured performance, need to be validated with a pathology-centric radiologist study before they can be considered relevant to clinical practice.* Please let us know if our proposed remedies fall short. We thank you again for your thorough review, continued discussion, and help in improving our submission.
I think the proposed limitations satisfy my primary concerns, and I have increased my score to 6.
Thank you for your response. I have read the other reviews and responses. I appreciate the discussion with reviewer N79n and the limitations added to the final version of the paper and further the additional results for the VarNet as a baseline, which shows that the quantitative performance of the approach is competitive with state-of-the-art end-to-end learning. I also acknowledge that the main goal of this work was to propose a fast and accurate posterior sampler and that the application of this sampler for uncertainty quantification is outside the scope of this paper. Hence, I increased the contribution socre from 2 to 3 and the overall score from 6 to 7.
Official Comment by Reviewer pyUh
The reviewer thank you the authors for addressing the reviewers’ concerns. After going through the results in the new adding materials and the discussions between other reviewers, I tend to accept this submission. However, in the context of Medical imaging and specific to MRI it’s unclear whether perceptual metrics are really relevant. It is clear that E2E-VarNet still works the best in terms of PSNR by 1.07 dB. In addition, seemingly the choice of $P$ may affect the final reconstruction results a lot, making the practical usage of this algorithm unclear. As a result, I can not increase my score as its current state. That is being said the review does not feel disappointed if this paper dose not got accepted eventually.
Thanks for your comments and continued discussion. **"In the context of Medical imaging and specific to MRI it's unclear whether perceptual metrics are really relevant"** - What matters in MRI is the ability of the radiologist to make the correct diagnosis. Studies like [A,B,C,D] have shown that well-chosen perceptual metrics correlate far better to radiologists perceptions than traditional metrics like PSNR. **"It is clear that E2E-VarNet still works the best in terms of PSNR by 1.07 dB"** - First, please see our previous comment about how PSNR is not a meaningful visual quality metric for MRI. - Second, the E2E-VarNet generates only a point estimate, not posterior samples. The subject of our paper is posterior sampling, not point estimation. Posterior sampling has many applications, only one of which is constructing point estimates. - Third, when used to construct point estimates, our approach beats the E2E-VarNet in *every* tested metric except PSNR. **"The choice of $P$ may affect the final reconstruction results a lot, making the practical usage of this algorithm unclear"** - Our paper is about fast and accurate posterior sampling, and $P$ is simply the number of samples that the user chooses to generate. - The effect of $P$ on reconstruction quality only manifests when creating point estimates, and the main subject of our paper is not point estimation. - Even if we do focus on point estimation, then the effect of $P$ is relatively clear. It is a direct way to navigate the perception-distortion tradeoff, which is a fundamental tradeoff in image reconstruction. Given a chosen image-quality metric, experiments like those in Table 2 clearly show how $P$ is best chosen. [A] S. Kastryulin, J. Zakirov, N. Pezzotti, and D.V. Dylov, “Image quality assessment for magnetic resonance imaging,” arXiv:2203.07809, 2022. [B] M.S. Treder, R. Codrai, and K.A. Tsvetanov, "Quality assessment of anatomical MRI images from generative adversarial networks: Human assessment and image quality metrics," Journal of Neuroscience Methods, vol. 374, 2022. [C] Mason, A., et al., 2020. Comparison of Objective Image Quality Metrics to Expert Radiologists’ Scoring of Diagnostic Quality of MR Images. IEEE TMI 39, 1064–1072 [D] H. Liu, W. Zhou, and R. Narayan. "Perceptual quality assessment of medical images." In Encyclopedia of Biomedical Engineering, pp. 588-596. 2019.
Thanks for taking the time to address my previous questions and concerns. After reading your feedback, the global rebuttal, and reviews from other reviewers and discussions, most of my concerns have been addressed and I do value the quality of this paper; consequently, I have raised my rating.
Decision
Accept (poster)