List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound imaging is widely used in clinical practice but is strongly affected by speckle noise, a signal-dependent pattern that blurs anatomical details and reduces diagnostic quality. Classical PDE-based despeckling methods provide interpretable solutions but rely on manually designed diffusion functions and parameter tuning, limiting adaptivity. In contrast, deep learning models achieve strong restoration performance but often lack interpretability and disregard the underlying PDE dynamics. In this paper, we propose a model-guided hybrid framework inspired by a recent hyperbolic–parabolic PDE scheme. Instead of analytically defining the diffusion function and solving an auxiliary equation, we approximate its key components using neural networks and learn the entire diffusion function in a data-driven manner. The diffusion evolution is reformulated as a learnable dynamical system, where a Mamba-based state space model, with high flexibility and efficiency in capturing complex and non-linear denoising dynamics, approximates the diffusion function and updates its components at each iteration. This allows the diffusion to adapt to local intensity and edge information while preserving the primary PDE structure. Extensive experiments on BUSI and AULI datasets show consistent improvements over classical diffusion methods and recent deep models in terms of PSNR, SSIM and PieAPP, effectively bridging model-based interpretability with data-driven flexibility for ultrasound despeckling. The source code is available at:
https://github.com/duypham01/Mamba-Diff-US.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5381_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/duypham01/Mamba-Diff-US
Link to the Dataset(s)
BUSI dataset: https://www.kaggle.com/datasets/sabahesaraki/breast-ultrasound-images-dataset
AULI dataset: https://www.kaggle.com/datasets/orvile/annotated-ultrasound-liver-images-dataset
BibTex
@InProceedings{PhaNha_Mamba_MICCAI2026,
author = { Pham, Nhat Duy AND Luong, Marie AND Kaaniche, Mounir AND Trinh, Dinh Hoan},
title = { { Mamba Based Anisotropic Diffusion Model for Speckle Noise Reduction in Ultrasound Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the challenge of speckle noise reduction in ultrasound imaging by proposing a Mamba-based anisotropic diffusion model. The core contribution lies in the integration of the Mamba state space model into a system of hyperbolic-parabolic partial differential equations (PDEs) to learn the diffusion functions. Empirical results demonstrate that the proposed method achieves noticeable improvements in terms of PSNR, SSIM, and PieAPP.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper focuses on a core challenge in medical ultrasound imaging, and the research direction is closely aligned with the current academic frontier, offering significant research value for solving practical problems in this field. 2.This work attempts to introduce Mamba, characterized by its linear complexity, into the field of ultrasound speckle noise suppression. Given that speckle noise in ultrasound images often exhibits complex spatial correlations, Mamba’s advantages in capturing long-range dependencies provide a new theoretical perspective for effectively distinguishing anatomical structures from coherent noise.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The core contribution lacks sufficient novelty, as it mainly combines existing methods without proposing new theoretical developments. This is the main weakness. 2.In the experimental results, the PSNR values at σ= 0.3 in Table 1 show that the proposed method is significantly inferior to the HCTSpeckle method. The authors must explicitly explain the cause of this performance gap. It should be clarified whether this shortfall stems from inherent limitations of the algorithm or is a deliberate design choice made to balance model complexity and precision. 3.Regarding the ablation study, categorizing the impact of diffusion steps as an “ablation” appears inappropriate. Ablation studies are typically intended to verify the contribution of specific modules or loss functions, whereas the number of diffusion steps is more accurately framed as a hyperparameter sensitivity analysis. I suggest moving this content to a “Sensitivity Analysis” section or redefining the scope of the ablation study to focus on the addition or removal of core structural components or algorithmic logic.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main reason is that he core contribution lacks sufficient novelty, as it mainly combines existing methods without proposing new theoretical developments.
- 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.
The response addresses several of the reviewers’ questions in a reasonable way. I especially appreciate that the authors clarified the interpretability claim, agreed that the diffusion-step analysis should be presented as a sensitivity analysis rather than a true ablation, and provided additional implementation details for reproducibility.
However, after carefully re-reading the paper and the rebuttal, I do not think the response sufficiently resolves my main concern, which is about the depth of the methodological contribution rather than presentation alone.
My primary concern was not whether the proposed method is different from prior work in a narrow descriptive sense, but whether it constitutes a sufficiently strong methodological advance for MICCAI. In this regard, the rebuttal mainly clarifies how the proposed framework differs from existing PDE-guided or direct-prediction approaches: it preserves an explicit iterative hyperbolic-parabolic PDE structure, decomposes the diffusion function into two meaningful components, and estimates them using Mamba-based models at each iteration. This clarification is helpful, but in my view it does not fundamentally change the core assessment that the work remains largely an incremental hybridization of existing ingredients rather than a substantial new theoretical or algorithmic development.
The rebuttal improves the framing of the contribution more than it strengthens the contribution itself. However, the method is still built on an existing PDE model, replaces manually designed terms with learned approximators, and instantiates these approximators with an existing MambaIRv2 backbone. While this is a reasonable and technically sound design, I am not yet convinced that it reaches the level of novelty expected for acceptance.
Regarding the performance gap on BUSI at (sigma = 0.3), I appreciate the authors’ clarification that this reflects a trade-off between pixel-wise fidelity and structural/perceptual preservation. I think this is a fair explanation. Likewise, I appreciate the acknowledgement that the diffusion-step experiment is more appropriately viewed as sensitivity analysis. These points improve the presentation and interpretation of the work.
However, these clarifications do not fully address the more central issue: the paper still lacks stronger evidence that the proposed design choices are necessary and methodologically compelling beyond being a reasonable engineering combination. In particular, the rebuttal acknowledges that additional comparisons or ablations, such as against simpler formulations or alternative design choices, would be interesting, but does not provide stronger support on this front. As a result, the main question of contribution strength remains.
Overall, I find the paper technically competent and the rebuttal professional. However, my concern was always primarily about methodological depth and novelty, and I do not think the rebuttal sufficiently changes that assessment. Therefore, after considering the rebuttal, I maintain my recommendation: Reject.
Review #2
- Please describe the contribution of the paper
Proposed a hybrid ultrasound despeckling framework that combines a hyperbolic–parabolic PDE formulation with learnable diffusion estimation using Mamba-based networks. Rather than directly predicting a denoised image, the method approximates key components of the diffusion control function within an iterative PDE-inspired update scheme. Specifically, two Mamba-based modules are used to estimate the gray-level and edge-related terms of the diffusion coefficient, aiming to preserve the interpretability and structure of classical anisotropic diffusion while introducing data-driven adaptability. The method is evaluated on BUSI and AULI ultrasound datasets with synthetic speckle corruption and is compared against several recent ultrasound despeckling baseline methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a relevant problem in ultrasound imaging, namely speckle reduction while preserving diagnostically meaningful structures. This is an important challenge for both visual interpretation and downstream automated analysis. 2.The overall idea is interesting and reasonably motivated. The authors attempt to bridge classical PDE-based despeckling, which is interpretable but rigid, with modern learned restoration models, which are flexible but often less physically grounded. The proposed approach is conceptually coherent and aligns with an active direction in model-guided medical image restoration. 3.The method is not a purely black-box denoiser. By embedding the learnable components within a diffusion-inspired iterative way, the paper retains some structural connection to established anisotropic diffusion models rather than completely discarding the PDE formulation. 4.Quantitatively, the proposed method is competitive across both BUSI and AULI and performs favorably against multiple recent baselines under the reported experimental setting. The gains are not dramatic in all cases, but the method is generally strong and consistent. 5.The paper includes ablation analysis, including one-vs-two Mamba approximators and the effect of the number of diffusion steps, which is helpful even though it is still limited.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The main weakness is that the experimental validation is not strong enough. The evaluation is limited to two relatively small 2D ultrasound datasets with synthetically added speckle noise. This setup is useful as an initial benchmark, but it is not sufficiently convincing for establishing clinical robustness or real-world value in ultrasound restoration. Performance on real acquisition noise, cross-device settings, or downstream tasks would make the paper much stronger. 2.The novelty is moderate. The manuscript itself positions the work against several recent hybrid or learned diffusion approaches, including learning diffusion functions for restoration and ultrasound diffusion-based denoising models. Relative to these, the main novelty seems to be the specific use of Mamba-based approximators and the decomposition of the diffusion term into two learned components. This is interesting, but currently feels more like an incremental methodological extension. 3.The ablation study is insufficient to support the central claims. In particular, the paper does not adequately isolate whether the gains come from the PDE-inspired formulation, the Mamba backbone, the two-branch diffusion decomposition, or simply from having a strong learned restoration model. Stronger controls would include Mamba vs CNN/UNet/Transformer with matched capacity, hyperbolic–parabolic vs purely parabolic learned diffusion, and PDE-guided vs direct image prediction baselines. 4.Important implementation details remain unclear or insufficiently emphasized for full reproducibility, including details of the numerical discretization, the role of the theta-scheme parameter, constraints on the learned diffusion maps, iteration-wise parameter sharing, and the exact practical realization of the PDE operator. 5.The paper does not report statistical significance testing, repeated-run variability, runtime/efficiency comparison, or patient-level split discussion. These omissions reduce confidence in the robustness and practical significance of the reported gains.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles an important ultrasound despeckling problem and proposes an interesting hybrid framework that combines a hyperbolic parabolic PDE formulation with learnable Mamba-based diffusion approximation. I appreciate the attempt to preserve model based structure instead of using a fully black-box denoiser, and the method is competitive on the reported BUSI and AULI benchmarks. The rebuttal helps to clarify parts of the motivation and implementation, and I acknowledge that the method is technically coherent. That said, my overall score remains below the acceptance threshold. The main reason is that the empirical validation is still not strong enough. The experiments are limited to two 2D ultrasound datasets with synthetic speckle corruption, and the ablation studies do not sufficiently isolate the effect of the PDE formulation, the Mamba backbone, or the claimed interpretability advantages. In addition, the novelty relative to recent learned diffusion or PDE-guided restoration works appears moderate rather than substantial. While the reported quantitative performance is good, the gains are generally modest and not supported by deeper robustness analysis, significance testing, or stronger real-world validation. Overall, I see this as a promising and well-motivated work, but in its current form it reads more as a good incremental contribution. The rebuttal improves clarity, but it does not fully overcome the concerns regarding experimental rigor, and strength of novelty.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The paper addresses a clinically relevant problem and proposes a conceptually coherent framework that its learnable hyperbolic–parabolic diffusion decomposition with iteration-wise Mamba-based estimation, going beyond simple CNN/UNet single-diffusion approximations. The method is quantitatively competitive and consistent across two benchmarks. The authors rebuttal clarifies the novelty claims, and commits to full reproducibility via code release and additional implementation details. On balance, the work represents a good incremental to moderate contribution.
Review #3
- Please describe the contribution of the paper
This paper proposes a physics-guided ultrasound despeckling framework that combines a hyperbolic–parabolic PDE formulation with learnable diffusion estimation. Instead of manually specifying the diffusion function and explicitly solving the auxiliary PDE variable, the method learns the two components of the diffusion coefficient with two MambaIRv2 networks and embeds them into the iterative PDE update. The authors evaluate the method on BUSI and AULI after adding synthetic speckle noise at three corruption levels, and report competitive PSNR, SSIM, and PieAPP against several recent ultrasound denoising baselines.
- 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 main technical idea is sensible and reasonably elegant. Rather than discarding the PDE prior, the method preserves the structure of the hyperbolic–parabolic model and only replaces the hand-designed diffusion terms with learned approximators. The factorized design, where one network models the gray-level term and another models the edge-related term, is more interpretable than a fully black-box direct image-to-image denoiser.
The experimental results are competitive. On BUSI and AULI, the method is usually best or near-best across the reported noise levels. The paper also includes a useful ablation comparing one-network vs two-network diffusion approximation and an analysis of the number of diffusion steps, showing that the factorized version performs better and that the iterative process stabilizes around 40 steps.
The paper is generally organized in a logical way, and the central idea is easy to follow from the formulation through the experiments.
- 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 evaluation setting is limited. The experiments are performed by resizing BUSI and AULI images to 224×224 and adding synthetic speckle using the paper’s noise model. There is no evaluation on raw real-noisy ultrasound acquisitions, no cross-device or cross-site robustness test, and no downstream-task assessment such as segmentation or diagnosis performance after denoising, even though the introduction explicitly motivates the method through downstream analysis quality. This makes it difficult to judge practical value.
The performance gains are not uniformly strong. The method is competitive overall, but the improvements over the best baseline are often small, and sometimes another method is better on an individual metric. For example, on BUSI at σ=0.3, HCTSpeckle has higher PSNR than the proposed method (28.81 vs 28.56), while the proposed method is better in SSIM and PieAPP. On AULI at σ=0.4, LAD-CNN has slightly higher PSNR (31.84 vs 31.82), while the proposed method has better SSIM and PieAPP. So the empirical story is “strong and competitive” rather than clearly dominant.
The ablation study is too narrow for the claims being made. At least include: hyperbolic–parabolic vs purely parabolic learned diffusion, effect of removing the PDE update and directly predicting the output, and statistical testing. The current ablations only test one-vs-two Mamba networks and number of diffusion steps on one dataset/noise level.
The paper also overstates interpretability. The PDE scaffold is interpretable, but once both diffusion components are produced by learned Mamba networks, interpretability is only partial unless the authors actually visualize or analyze the learned coefficient maps and show correspondence to intensity structure or edges. That evidence is currently missing.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The submission is well motivated, presents a technically sound hybrid design, and demonstrates competitive performance. While the observed improvements are generally incremental rather than transformative, the study provides a solid foundation for future development. The evaluation is currently focused on synthetic-noise benchmarks, and the work could be further strengthened by including real clinical robustness assessments and downstream task validation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors clarified the intended novelty over prior PDE-guided despeckling methods, the use of synthetic speckle benchmarks, the partial nature of the interpretability claim, and several implementation details. Limitations remain: particularly the lack of real-noisy ultrasound validation, cross-device robustness analysis, downstream-task evaluation, and broader ablations. The paper addresses an important ultrasound restoration problem, presents a technically coherent hybrid framework that combines explicit hyperbolic–parabolic PDE evolution with learned Mamba-based diffusion components, and reports competitive results on BUSI and AULI. The contribution is incremental rather than transformative, but it is well motivated.
Author Feedback
We sincerely thank the reviewers and ACs for their valuable comments and constructive feedback. We appreciated their overall positive comments and addressed below their concerns. 1-Methodological novelty (Rev. 1, Rev. 2, AC) Our contribution is not a simple combination of existing modules. Prior PDE-guided despeckling works either: (i) use fixed analytical diffusion functions with manually tuned parameters, or (ii) approximate a single parabolic diffusion process using CNN/UNet architectures. In contrast, our work introduces a learnable hyperbolic–parabolic diffusion framework in which, (a) the PDE evolution itself remains explicit and iterative, (b) the diffusion function is decomposed into two physically meaningful components corresponding to gray-level and edge-detector diffusion terms, and (c) these components are adaptively estimated at every iteration using Mamba-based state-space models. Unlike prior direct image-prediction approaches, the denoised image is generated through iterative PDE evolution governed by learned diffusion dynamics. 2-Experimental setting and practical validation (Rev. 2, 4, AC) We agree that validation on real-noisy data or downstream tasks could further strengthen the experiments. However, the current setting follows the standard benchmark protocol adopted by prior works (e.g., USNet, LAD-CNN, HCTSpeckle, and SRAD-ResUNet), which also rely on synthetic speckle generation due to the absence of paired clean/noisy data. Our goal in this work is to provide a controlled and reproducible comparison against recent state-of-the-art methods under identical degradation settings. 3-Interpretability claim (Rev. 4, AC) Let us clarify that interpretability is partial rather than absolute. Our claim is mainly that the proposed design preserves the PDE structure and maintains physically meaningful diffusion components instead of directly reconstructing the denoised image (i.e. instead of using a fully black-box denoiser). The learned terms still correspond to gray-level and edge-detector modulation within the diffusion process. We will slightly revise the wording to avoid overstating interpretability and restrict its use to the model design. 4-Ablation and sensitivity analysis (Rev. 1, 2, 4) First, we agree that the impact of diffusion steps should appear as a sensitivity analysis. Regarding the suggested ablations (e.g., purely parabolic formulation, direct prediction baseline, alternative backbones), we acknowledge their interest. However, due to space limitation, we proposed to illustrate the effect of decomposing the diffusion function into two learned components versus a single learned approximation, directly supporting the proposed design. 5-Performance at some specific noise levels (Rev. 1, 4) On BUSI (σ=0.3), HCTSpeckle achieves slightly higher PSNR (28.81 vs 28.56 dB), while our method achieves better SSIM and PieAPP, indicating improved structural preservation and perceptual quality. Similarly, on AULI (σ=0.4), LAD-CNN obtains marginally higher PSNR while our method performs better on perceptual/structural metrics. This reflects the trade-off between aggressive smoothing optimized for pixel-wise fidelity and PDE-guided diffusion designed to better preserve structures and perceptual consistency. Overall, as mentioned by Rev. 4, we agree that the proposed method is “strong and competitive” rather than uniformly dominant. 6-Implementation details and reproducibility (Rev. 2) Let us recall that the main details are provided in Sections 2 and 3.2 (e.g., PDE formulation, iterative update rule, weighted θ-scheme discretization, diffusion-step setting, MambaIRv2 architecture). Moreover, further implementation details could be easily included in the final version (e.g., θ set to 0.4, learning diffusion maps constrained to the range [0,1], and parameter-sharing across iterations for the two Mamba networks). For reproducibility purposes, we are also providing a link to the source code of the proposed framework.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers acknowledge that the proposed method is competitive and addresses a relevant problem. However, several concerns remain regarding the significance and validation of the contribution. In particular, the methodological novelty appears moderate relative to existing approaches, and the current experimental setup, based on synthetic noise and limited datasets, does not yet provide sufficient evidence for robustness or practical impact. In addition, some claims (e.g., interpretability and effectiveness of specific components) are not fully supported by the current analysis.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After considering the rebuttal and the updated reviews, I find that the major concerns have been reasonably addressed. While Reviewer 1 remains unconvinced about the level of methodological novelty, the rebuttal clarifies the contribution and the remaining concerns are primarily about contribution strength rather than technical validity. Given the positive assessments from Reviewers 2 and 3, the technical soundness of the method, and its relevance to the community, I support acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal, the opinions are mixed but overall support acceptance: two reviewers recommend Accept, while one reviewer remains concerned about methodological novelty and depth. The authors provided reasonable clarifications on the novelty claim, the use of synthetic benchmarks, the scope of interpretability, and implementation/reproducibility details. I agree that the contribution is incremental rather than transformative, and that stronger real-noisy validation and broader ablations would improve the work. However, these limitations do not appear fatal. Given the technical coherence, competitive results, and positive post-rebuttal assessments from the majority of reviewers, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a technically sound approach for ultrasound image despeckling using Mamba-based anisotropic diffusion model. Although the method is not strictly novel and the evaluation is also restricted to synthetic speckle, it should be of interest to the community working in the medical image enhancement domain.