Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Fluorescence microscopy is an indispensable tool for biological discovery but image quality is constrained by desired spatial and temporal resolution, sample sensitivity, and other factors. Computational denoising methods can bypass imaging constraints and improve signal-to-noise ratio in images. However, current state of the art methods are commonly trained in a supervised manner, requiring paired noisy and clean images, limiting their application across diverse datasets. An alternative class of denoising models can be trained in a self-supervised manner, assuming independent noise across samples but are unable to generalize from available unpaired clean images. A method that can be trained without paired data and can use information from available unpaired high-quality images would address both weaknesses. Here, we present Baikal, a first attempt to formulate such a framework using Denoising Diffusion Probabilistic Models (DDPM) for fluorescence microscopy images. We first train a DDPM backbone in an unconditional manner to learn generative priors over complex morphologies in microscopy images, we can then apply various conditioning strategies to sample from the trained model and propose optimal strategy to denoise the desired image. Extensive quantitative comparisons demonstrate better performance of Baikal over state of the art self-supervised methods across multiple datasets. We highlight the advantage of generative priors learnt by DDPMs in denoising complex Flywing morphologies where other methods fail. Overall, our DDPM based denoising framework presents a new class of denoising method for fluorescence microscopy datasets that achieve good performance without collection of paired high-quality images.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3885_paper.pdf

SharedIt Link: https://rdcu.be/dV5BG

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72104-5_12

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/3885_supp.pdf

Link to the Code Repository

https://github.com/scelesticsiva/denoising

Link to the Dataset(s)

https://publications.mpi-cbg.de/publications-sites/7207/

BibTex

@InProceedings{Cha_Baikal_MICCAI2024,
        author = { Chaudhary, Shivesh and Sankarapandian, Sivaramakrishnan and Sooknah, Matt and Pai, Joy and McCue, Caroline and Chen, Zhenghao and Xu, Jun},
        title = { { Baikal: Unpaired Denoising of Fluorescence Microscopy Images using Diffusion Models } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15007},
        month = {October},
        page = {119 -- 129}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    In this article, the author applied DDPM to denoise fluorescence images without needing paired training data. The author ingeniously applied the idea of inpainting to denoising and explores the fusion of the original image in the reverse process around three sampling methods and conducted a series of experiments to validate the superiority of the denoising effectiveness.

  • Please list the main strengths of the paper; you should write about 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.

    Self-supervised denoising methods assume pixel-wise independent noise, which may lead to worse performance under structured noise. Baikal utilizes DDPM to learn data distributions from clean images and applies them for denoising, enabling the model to generate denoised results that better match reality. In the reverse process of generating clean images, the idea of inpainting is adopted to integrate the original image information. As expected, this approach has achieved performance second only to fully supervised denoising models. The authors evaluated the impact of incorporating self-supervised methods on conditional samplers by feeding the results processed by N2V into the reverse process. Investigating whether the synthesis of two denoising methods can lead to further gains is meaningful. The authors also conducted very detailed ablation experiments, which included the selection of different sampling strategies and hyperparameters in sampling, ultimately determining the optimal implementation of the algorithm.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    In the experiment, it is unreasonable to only compare with the self-supervised algorithm N2V. But as far as I know, N2V is already an approach from several years ago. In recent years, there have been many self-supervised denoising methods [1, 2] that far exceed its performance, yet the authors neither introduced them in the introduction nor compared them in the experiments.

    To fully leverage clean images and demonstrate the superiority of the DDPM framework, I suggest that the authors compare their method with denoising approaches that can utilize clean image information, such as those based on generative adversarial networks.

    Furthermore, I think there is limited experimental support for the third contribution of the paper. The author did not test the impact of multiple self-supervised denoising algorithms on conditional samplers. Therefore, it is hard to verify whether it is due to the design of the N2V algorithm itself or if all self-supervised algorithms are not suitable for denoising fluorescence microscopy images.

    [1] Li J, Zhang Z, Liu X, et al. Spatially adaptive self-supervised learning for real-world image denoising[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 9914-9924. [2] Zhang Y, Li D, Law K L, et al. Idr: Self-supervised image denoising via iterative data refinement[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 2098-2107.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    There are relatively few comparative methods in the experiments. I suggest that the authors further supplement the experiments to validate Baikal’s superiority over the latest self-supervised denoising methods. Additionally, it is recommended that the authors conduct detailed research on current self-supervised algorithms in the review, including some methods that model structural noise.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    In summary, I believe the work of employing diffusion models for denoising fluorescence microscopy images exhibits a certain level of innovation. During model training, the availability of clean image data is more feasible than paired image data, thus providing broader application scenarios. Furthermore, the writing is clear and the organization is logical. However, the “Introduction” section lacks a comprehensive survey of self-supervised denoising methods and lacks a compilation of relevant works on using DDPM for denoising. Although the experimental section of the paper presents a considerable amount of visualization, the method comparison is relatively limited. There is no comparison with the current state-of-the-art self-supervised denoising methods.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • [Post rebuttal] Please justify your decision

    The author clarifies that the main contribution of the article is to evaluate and suggest the best conditional samplers from the inpainting DDPM literature for the fluorescence microscopy denoising task. I acknowledge this motivation, but I believe that experimental comparisons with SOTA methods are necessary for the following reasons:

    (1) The denoising process of diffusion models often requires multi-step computations, which brings a greater computational burden. The manuscript needs to demonstrate through comparison with the other methods whether the introduction of denoising with diffusion models is necessary or whether it offers advantages over traditional methods.

    (2) The paper explores the optimal implementation of conditional samplers from previous literature, without proposing a new method, and I think the innovation is not particularly strong.

    I appreciate the additional experiments provided by the reviewer. However, these cannot serve as the basis for determining the results.



Review #2

  • Please describe the contribution of the paper

    This paper applies diffusion models to denoise fluorescence microscopy images. It experimented with multiple reverse diffusion methods during inference stage and demonstrated DDPM-based denoising method is able to outperform existing blind image denoising methods (N2V, CARE) .

  • Please list the main strengths of the paper; you should write about 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.

    As an application paper, it provides comprehensive details on methods to generate denoised images during the inference stage(forward-backward, mixing, and repainting). Additionally, it discusses when certain conditional sampling techniques are likely to outperform others and the selection of optimal hyperparameters through extensive experimentation, offering valuable insights into the proper utilization of DDPMs for fluorescence microscopy image denoising.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    DDPMs, the paper introduces novelties in the inference stage through modified mixing and repainting techniques[1]. Furthermore, the application of DDPMs to fluorescence microscopy image denoising is a novel contribution. [1] A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte and L. Van Gool, “RePaint: Inpainting using Denoising Diffusion Probabilistic Models,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The paper used an open source code so it should be reproducible

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    In addition to the main weakness, it will be helpful to have more clarification on N2V related experiments - why is N2V output used as input to the reverse process? Output from N2V is already denoised and is the paper trying to do some additional denoising on the already denoised data?

    It will also be helpful for the authors to provide insights on training time and evaluation time for DDPM, size of input, this will help readers understand the clinical applicability of the proposed method as DDPM typically has limitations on image size and suffered from long training/inference time.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the paper is a good application paper (with minimal algorithm novelty), it will need to address some concerns on actual applicability.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    Thank you for the response. The author addressed most of my questions and from the paper offers useful insights in DDPM application to microscopy images. It will be useful to provide more info on timestep “t” in diffusion model as well in revision.



Review #3

  • Please describe the contribution of the paper

    The manuscript titled, “Baikal: Unpaired Denoising of Fluorescence Microscopy Images using Diffusion Models” demonstrates a methodology that utilizes Denoising Diffusion Probabilistic method trained on unpaired data for denoising fluorescence images.

  • Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper’s main strengths lie in the novel implementation of the DDPM method using unpaired clean imaging datasets. The manuscript also nicely discusses other denoising methods that are either paired or self-supervised denoising methods and indicates the disadvantage of the prior methods in working with unpaired imaging datasets.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    Although the manuscript discusses the novel application of denoising using DDPM in fluorescence images, it does not discuss how the output of denoising compares with deconvolution tools present in typical confocal fluorescence microscopes like Deltavision. It is hard to understand from the manuscript if this method is needed for denoising images collected using fluorescence microscopy. Additionally, without paired ground truth datasets for testing, it is difficult to confirm the improvement in denoising.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Comments to authors: The manuscript titled, “Baikal: Unpaired Denoising of Fluorescence Microscopy Images using Diffusion Models” demonstrates a methodology that utilizes the Denoising Diffusion Probabilistic method trained on unpaired data for denoising fluorescence images. The paper’s main strengths lie in the novel implementation of the DDPM method using unpaired clean imaging datasets. The manuscript also nicely discusses other denoising methods that are either paired or self-supervised denoising methods and indicates the disadvantage of the prior methods in working with unpaired imaging datasets. Although the manuscript discusses the novel application of denoising using DDPM in fluorescence images, it does not discuss how the output of denoising compares with deconvolution tools present in typical confocal fluorescence microscopes like Deltavision. It is hard to understand from the manuscript if this method is needed for denoising images collected using fluorescence microscopy. Additionally, without paired ground truth datasets for testing, it is difficult to confirm the improvement in denoising. The scope of this work seems limited, as the manuscript doesn’t highlight other applications in biomedical imaging. Altogether, this work could benefit from (1) comparing the denoising outputs for fluorescence images collected at different resolutions, (2) a description of other applications to highlight biomedical relevance, and if possible, the need and superior performance of this method in denoising fluorescence imaging.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Major factors to justify my recommendation include: Limited scope

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    The authors have sufficiently addressed all my comments, I recommend accepting the manuscript in its current form.



Review #4

  • Please describe the contribution of the paper

    The authors propose a new unsupervised deep learning approach for denoising of 2D fluorescence microscopy images using diffusion models. They address the issue of maintaining the structural information of the noisy image after denoising by combining the result of the different denoising steps in the diffusion model during inference. This way of conditioning the final result improves the accuracy of the method. In their work they test the proposed method on three different datasets and they compare themselves with state-of-the-art and well-established methods in the field, showing the expected superiority of diffusion models.

  • Please list the main strengths of the paper; you should write about 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.

    Image denoising or restoration in fluorescence microscopy is an extremely useful approach to push live-cell imaging. Due to the challenges, or even impossibility, associated to acquiring paired field of views of the data, unsupervised methods are highly demanded in the field. The authors in this case take advantage of the very new diffusion models, which are outperforming existing methods in plenty of images modalities but are yet unexploited and unexplored in the field of microscopy. Moreover, diffusion models are usually not conditioned, which is not that useful in microscopy or even recommended for image driven discovery. For this, they addapt the denoising workflow of the classical diffusion models in a simple but still reasonable manner, to support the preservation of the structural information of the input image in the output. For this, they use the “Repainting” strategy from image painting, which I wonder how much it could be pushed in this type of generative approaches to prevent hallucinations.

    While expected, the proposed method outperforms the accuracy of established methods in the field, both supervised and self-supervised in three different datasets.

    The method is in general, well described, the figures are quite appropriate and support the text, and the authors explain objectively the different limitations of the proposed approach.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    In general the work presented here is very much needed in the field and knowing that it is using generative AI, one of the main weaknesses to highlight is the use of only pixel-level accuracy metrics (SSIM, MSE and PSNR). I wonder wether the authors could have benchmarked the proposal with perceptual metrics (e.g., LPIPS, NIQE) as well, which could potentially show the strength of the method.

    The authors claim that there are no works using diffusion models for fluorescence microscopy but I would suggest reviewing the literature again. While there are not directly addressing conditional denoising, I would smooth the statements in the text. Here are some works:

    • Denoising Diffusion Probabilistic Models for Generation of Realistic Fully-Annotated Microscopy Image Data Sets https://doi.org/10.48550/arXiv.2301.10227
    • Microscopy image reconstruction with physics-informed denoising diffusion probabilistic model, https://doi.org/10.48550/arXiv.2306.02929
    • DiffuseIR: Diffusion Models for Isotropic Reconstruction of 3D Microscopic Images, https://doi.org/10.1007/978-3-031-43999-5_31 (MICCAI 2023)

    The test dataset used in this work was excluded due to the “lack of normalisation”. What is the type of normalisation done in the training data? Can it be replicated?

    Relationship between the unsupervised/non-paired training and the real data: First, It is not clear how the training is conducted. In Figure 1 it sais: “using only unpaired high quality images”. Do you refer using only high quality images? Then, how does the noise variability or the noise itself in the entire data affect to the result? Specially the one in the unseen data, as diffusion models usually simulate gaussian noise but not the common Poisson noise in fluorescence microscopy. Also, would it be an issue if the data displayed noise and photobleaching in a random way in the “noisy” data? For example, if one would like to use the current method to denois a live-cell time-lapse imaging experiment.

  • Please rate the clarity and organization of this paper

    Very Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The authors did not directly provide their code or access to their code but they mention the source code for the diffusion model, not for the Baikal approach. The datasets used here are all publicly available.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    • The denoising network is described as mu_theta. Can you please define theta in the text?

    • In the description of the methods, the authors mention several times that the input y_t is independent of DDPM output. In what sense? Isn’t it conditioned and dependent structurally? Independence is used along the text in some other occasions. I would suggest further explaining in which sense the authors consider the samples being independent.

    • It is not clear in the text what is exactly the stop time step and how it affects to the mixed and repaint strategies. Same for the starting time step. Likewise, in the description of the method, the authors do not mention any weight schedule, that latter is shown in Figure 2

    • Table 1: I strongly recommend the authors displaying the distribution behind these numbers at least in the supplementary material, or providing the standard deviation.

    • Page 2: “Our method, which we term Baikal” –> Our method, which we termed Baikal

    • Page 4: “Aditionally, information…” –> Aditionally, the information…

    • Page 4: “A disadvantage of this sampler is that, as we re-introduce noisy image at every time step by taking weighted combination, it replaces denoised pixels by noisy pixels reverting denoising.” –> A disadvantage of this sampler is that it replaces denoised pixels by noisy pixels reverting denoising, as we re-introduce noisy image at every time step by taking weighted combination.

    • Page 6: “Notably, our proposed approach is more generalizable during training time.” What do you mean?

    • Figure 1: In general it is complete but to be more synchronised with the text, I would consider including “x” and “y” respectively as it is used in the technical description of the forward and reverse steps, as well as the mixed and repaint strategies.

  • 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

    Accept — should be accepted, independent of rebuttal (5)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The work presented by the authors is very much needed in the field as diffusion models are already being used in plenty of applications, and microscopy is yet one more. Also, the authors provide a method to get conditional diffusion models that seems to work well.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    I keep the same score as in the first review



Author Feedback

R1:1) The main aim of the paper is NOT to develop a SOTA denoising method, but to evaluate and suggest the best conditional samplers from the inpainting DDPM literature for the fluorescence microscopy denoising task. Therefore to show the relative performance of DDPMs we compared against two widely used methods N2V and CARE. However, upon suggestion by the reviewer, we compared the accuracy of our method with two additional methods namely SSID and IDR. Here we present SSIM metric comparisons for all datasets with methods mentioned in parentheses. Planaria - (0.270, 0.309, 0.296, 0.386)(SSID, IDR, N2V, Ours). Tribolium - (0.258, 0.301, 0.208, 0.458)(SSID, IDR, N2V, Ours). Flywing - (0.221, 0.163, 0.174, 0.290) (SSID, IDR, N2V, Ours). Indeed SSID and IDR performed better than N2V specially in Tribolium data that has structural correlated noise present in images. Still our method outperforms all these methods. 2) We will change the wording in the third contribution to reflect that we evaluated the support of N2V predictions for DDPMs in denoising (and not self-supervised methods in general). 3) Since the goal of this work is the first demonstration of DDPM’s efficacy for this task, we will leave comparisons with GANs as future work. We will also make sure to include a more comprehensive literature survey in the final manuscript. R3: We thank the reviewer for the positive feedback. Due to space limitations we chose to respond to major concerns, but will incorporate suggested clarifications in the final manuscript. R4: We thank the reviewer for positive comments however we do not see any weakness described in the main weakness question. 1) We use N2V outputs to condition DDPMs to test the hypothesis that initial denoising by N2V can further improve denoising during the generative process. However our results demonstrate that DDPMs conditioning on noisy images are better than conditioning on N2V predictions. 2) We split all datasets in a 80:10:10 ratio for train, eval and test sets..Each z-plane is treated as an individual image during training, resulting in 229120-Planaria, 198400-Tribolium, and 711200-Flywing train images. For evaluation, each z-plane is denoised individually, but max-projected for SSIM, MSE, and PSNR calculations, consistent with previous methods.Thus we have 1790-Planaria 1550-Tribolium and 1778-Flywing eval/test images. Size of input - 64x64 Training time: ~24 to 48hrs on a V100GPU, Evaluation time: <1second/image. 3) Rather than a new algorithm, this paper focuses on the novel application of DDPMs to denoise fluorescence microscopy images, acknowledging the unique noise characteristics that differentiate them from other imaging modalities. R5: We appreciate the positive feedback. 1) We would like to make a distinction between deconvolution and denoising. While deconvolution enhances spatial resolution, our method focuses on denoising fluorescence microscopy images, vital for removing photon and camera readout noise caused by low laser powers or brief exposures. 2) Denoising has several important applications such as preserving light-sensitive samples, reducing motion blur, neuron activity extraction [1,2], segmentation[3,4] and tracking, thus receiving considerable attention in the fluorescence microscopy field and therefore has broad applications. 3)We would like to clarify the training and testing procedure. We train the DDPM backbone using high quality clean images only from the train set. During testing/evaluation we DO have paired noisy and high quality images. Thus we denoise images using our DDPM method and evaluate the accuracy using high quality images in test/eval sets. 4)We were not able to compare against Deltavision as it is not open-source. However, we further demonstrate better performance of our method compared to 2 additional SOTA self-supervised denoising methods (see R1). [1]Buchholz,et al.ECCV,2020 [2]Stringer,et al.bioRxiv,2024 [3]Li,et al.Nat Methods,2021 [4]Lecoq,et al.Nat Methods,2021



Meta-Review

Meta-review #1

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The manuscript presents a novel approach to denoising fluorescence microscopy images using Denoising Diffusion Probabilistic Models (DDPMs) without requiring paired training data. By leveraging generative prior from paired clean images, the authors address the limitations of current supervised and self-supervised methods.

    The reviewers collectively acknowledged the innovative application of DDPMs for denoising fluorescence microscopy images. They appreciated the clear writing, logical organization, and comprehensive evaluation of conditional sampling strategies. The paper demonstrates significant potential for practical applications in biomedical imaging. The authors have addressed the major concerns raised, demonstrating the robustness and applicability of their method.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    The manuscript presents a novel approach to denoising fluorescence microscopy images using Denoising Diffusion Probabilistic Models (DDPMs) without requiring paired training data. By leveraging generative prior from paired clean images, the authors address the limitations of current supervised and self-supervised methods.

    The reviewers collectively acknowledged the innovative application of DDPMs for denoising fluorescence microscopy images. They appreciated the clear writing, logical organization, and comprehensive evaluation of conditional sampling strategies. The paper demonstrates significant potential for practical applications in biomedical imaging. The authors have addressed the major concerns raised, demonstrating the robustness and applicability of their method.



back to top