List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Sanqian Li, Risa Higashita, Huazhu Fu, Heng Li, Jingxuan Niu, Jiang Liu
Abstract
Anterior segment optical coherence tomography (AS-OCT) is a non-invasive imaging technique that is highly valuable for ophthalmic diagnosis. However, speckles in AS-OCT images can often degrade the image quality and affect clinical analysis. As a result, removing speckles in AS-OCT images can greatly benefit automatic ophthalmology analysis. Unfortunately, challenges still exist in deploying effective AS-OCT image denoising algorithms, including collecting sufficient paired training data and the requirement to preserve consistent content in medical images.
To address these practical issues, we propose an unsupervised AS-OCT despeckling algorithm via Content Preserving Diffusion Model (CPDM) with statistical knowledge.
At the training stage, a Markov chain transforms clean images to white Gaussian noise by repeatedly adding random noise and removes the predicted noise in a reverse procedure. At the inference stage, we first analyze the statistical distribution of speckles and convert it into a Gaussian distribution, aiming to match the fast truncated reverse diffusion process. We then explore the posterior distribution of observed images as a fidelity term to ensure content consistency in the iterative procedure.
Our experimental results show that CPDM significantly improves image quality compared to competitive methods. Furthermore, we validate the benefits of CPDM for subsequent clinical analysis, including ciliary muscle (CM) segmentation and scleral spur (SS) localization.
Link to paper
DOI: https://doi.org/10.1007/978-3-031-43990-2_62
SharedIt: https://rdcu.be/dnwMm
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a content-preserving diffusion model to denoise speckle noise in AS-OCT 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.
-It applies diffuse model to OCT denoising -It does not require paired training. -The method is validated in two dataset
- 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.
- Noise model. The speckle noise in OCT is widely considered to be with Rayleigh distribution (Orly Liba et al, 2017 Nature Comm). It is not clear how a Rayleigh distribution could be converted to Gaussian distribution. Simply using central limit theorem to model a noise distribution to Gaussian is not convincing.
- Diffusion model itself is not novel in denoising.
- The experimental setup and results evaluation is arbitrary. It is not clear why segmentation performance is evaluated in CM-Casia dataset but not in AS-Casia dataset. Similarly, the methods from which Fig. 3 is generated does not agree with methods from which Fig. 4 is generated.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
This study is not based on public dataset. The code is not open to public either.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- It will be great to make it consistent on methods shown in both Fig. 3 and Fig. 4.
- The noise model needs to be revisited and carefully evaluated.
- As the speckle noise is analyzed, it is important to add scale bar in Fig. 3 and Fig. 4. -It will be great elaborate how preserved content benefit clinical intervention.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
4
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The noise model in not convincing and the experiments are not well organized.
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
The authors made changes to partially address the concerns raised in the first round review.
Review #2
- Please describe the contribution of the paper
In this paper, a Content-Preserving diffusion model for unsupervised AS-OCT image despeckling is proposed. The method leverage the Gaussian assumption of diffusion models by converting specking through a logarithmic function. Truncated reverse diffusion procedure with data fidelity term was also introduced to effectively remove speckle noise.
- 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 method leverage the Gaussian assumption of diffusion models by converting specking through a logarithmic function which is novel for diffusion based denoising models.
- Detailed evaluation on downstream tasks, CM segmentation and SS localization task.
- 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.
- Probably the regions selected for calculating ENL should be reconsidered. ROI for ENL calculation should be homogeneous region not just background regions with low intensity. The reported ENL of CPDM is too high to be true.
- The process of calculating iteration number from estimated noise level is not well described in the paper which is the most important part of truncated reverse diffusion.
- The noise level estimation method for CPDM is [14], however for NLM, ANLM, WBM3D, and WK-SVD, it was conducted by [19]. This could make the comparison unfair.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
The proposed needs more implementation detail (noise level and iteration number estimation) to reproduce.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- Better ROI for calculating ENL.
- More detail description of iteration number estimation.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall technical novelty with some weakness in the method description and experiments set up.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes an unsupervised AS-OCT despeckling algorithm based on Content Preserving Diffusion Model (CPDM). The authors perform denoising by a DDPM and explore the posterior distribution of observed images as a fidelity term to ensure content consistency in the iterative procedure. The experimental results demonstrate the superiority of CPDM despeckling. In addition, the authors experimentally validate the effectiveness of CPDM for subsequent clinical analysis.
- 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 is clearly written and conducts a thorough literature study. The result section shows good improvement in despeckle performance on the AS-OCT. The authors also demonstrated the superiority of CPDM for speckle removal by completing the ciliary muscle (CM) segmentation and scleral spur (SS) localization task.
- 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.
- This paper is based on unsupervised. However, in the actual training phase, it is necessary to use clean images x_0 to train DDPM. This may not be considered as ‘unsupervised’ in the strict sense. Please give a better explanation.
- Please provide the original resolution size of the image before resizing.
- The speckle noise is defined as multiplicative noise in the paper. Can the authors give more relevant explanation?
- Most references lack information such as page and volume number.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
Might be difficult due to unavailability of dataset and code.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
Comments mentioned clearly in the strength and the weakness section of the papers.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
7
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The relevant weakness and comments are mentioned before but those are mostly minor and can be addressed quickly.
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
6
- [Post rebuttal] Please justify your decision
After referring to other reviewers’ comments and reading the authors’ responses, the reviewer revises the rating.
Review #4
- Please describe the contribution of the paper
The authors propose a despeckling algorithm of anterior-segment OCT (AS-OCT) based on a diffusion model with content preserving regularization (fidelity term) during the iterative denoising process.
- 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.
-
Considers that speckle noise violate the additive Gaussian assumption, and thus a log transform needs to be applied to fulfill the assumption.
-
Good evaluation of algorithm performance on two datasets, comparison with state-of-the-art denoising algorithms, the improvement of segmentation and localization performance on denoised images, and ablation studies showing the benefit of properly model speckle noise and the use of the fidelity term as regularization.
-
- 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.
-
Paper and in particular the method or algorithm is difficult to understand.
-
It is difficult to assess the contribution of the paper, in particular with regard to the method.
-
No validation set. It is unclear how hyperparameters (regularization lambda) has been determined.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
No code provided, no data provided. The algorithm is not well described. Paper cannot be easily reproduced.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
- It is not clear what the contribution of the paper exactly is, and how it differs to referenced methods (eg [13], [14]). Emphasize your contribution in the introduction.
Minor:
Typo: Page 8, OODM[15]->ODDM[15]
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Whereas on the one hand the evaluation is extensive and the despeckling seem to work well, it is very difficult to understand how the method works and how it can be applied. It is not clear what parts of the algorithm is novel and what part is derived from referenced work.
- Reviewer confidence
Somewhat confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. 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.
After reviewing the work, the three reviewers acknowledge its merit but raise few issues that should be addressed. First, two reviewers have raised issues about the statistics of speckle noise model in the submission. There are also issues with the validation setup (i.e. missing some important experiments, RIO selections, selection of the iteration number, potentially unfair comparisons).
Author Feedback
We’re grateful for the reviewers’ valuable comments. We are pleased that the reviewers recognized our contributions. Responses to major comments are provided below, but the space is insufficient to cover more detailed ones.
1.Noise model(R1) R: We checked the noise model is formally presented according to the previous study [16-18]. The speckle noise has a strong correlation with the signal in OCT system and is not independent. It multiplicative with the real signal[1-2]. The speckles are approximately transformed into additive Gaussian one by analyzing the statistical distribution [1-2]. We will add the log-transform distribution of speckles to improve its convincing while not only cite the central limit theorem. [1]. “Statistical Models of Signal and Noise and Fundamental Limits of Segmentation Accuracy in Retinal Optical Coherence Tomography.” TMI 2017. [2]. “Automatic Recovery of the Optic Nervehead Geometry in Optical Coherence Tomography.” TMI2007. 2.Contribution (R1&4) R: Different from the previous DDPM [13-15], our CPDM is specially designed for unsupervised image despeckling with content preserving. Firstly, it is impractical to directly apply the DDPM into the AS-OCT despeckling due to the multiplicative speckle noise is incompatible with the diffusion process, which adding the Gaussian repeatedly. We transform the speckles into the additive one by logarithm for fitting the diffusion process. Secondly, unsupervised DDPM [14-15] can easily miss inherent contents that are vital for clinic. So we design a fidelity, derived by the posteriori distribution of speckled images, into the iterative denoising process to achieve the content consistency. 3.Segmentation on AS-Casia dataset(R1) R: The images in AS-Casia dataset have no the segmentation label and we will add the related work in the future. 4.It is inconsistent on methods shown in both Figs.3-4 (R1) R: We only display the best six results among all methods due to the limited pages, and the results rank is different on the two datasets. 5.How preserved content benefit clinical intervention(R1) R: Structural content can benefit the clinical intervention in AS-OCT [1-2], and we validated the content preservation can improve the clinical auto-analysis in ablation study. The SS localization is critical for measuring anterior chamber angle and CM segmentation is benefit for myopia and presbyopia study[1]. [1].“Anterior Segment Optical Coherence Tomography.” Progress in Retinal and Eye Research, 2018. [2].“Attention to Region: Region-Based Integration-and-Recalibration Networks for Nuclear Cataract Classification Using AS-OCT Images.” MIA2022. 6.Better ROI for ENL(R2) R: We manually selected ROIs shown in the appendix, following the existing work in [8]. We changed the ROIs to calculate ENL, and CPDM always extremely surpass other competing methods (the results also have huge difference among different methods in [8]), we certainly checked the results and not chose opportunistically. Besides, we will publish the despeckling results for readers to calculate the ENLs with the ROIs that they are interested. 7.Iteration number evaluation; Noise level estimation is inconsistent(R2) R: Due to the limited pages and iteration number estimation is not our novelty, the details can refer to [14] and iteration number was evaluated as T=4 in the experiments. The noise level estimation in [14] also adopt the way in [19], we directly cite [14] due to further adopting iteration number estimation in [14].
- Unsupervised(R3) R: The unpaired data learning is considered as unsupervised [9-10], the CPDM further relieve the data requirement that only utilizing the clean data.
- Hyperparameters(R4) R: we analyzed the hyperparameters and set λ=μ⁄2M=0.2 in the experiment.
- Unavailability of dataset and code. R: Sorry for unable to disclose the data/code due to the confidentiality requirements of partner companies. If you’re interested, we’re happy to discuss it with you.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The authors have responded satisfactorily to the comments raised by the reviewers and my original comments. I recommend accepting the paper now.
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The authors have addressed the concerns raised by the reviewers and this paper got four positive ratings. I recommend accept.
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
This paper presents an innovative, unsupervised Anterior Segment Optical Coherence Tomography (AS-OCT) despeckling algorithm that’s predicated on a Content Preserving Diffusion Model (CPDM). The proposed methodology involves applying denoising through a Denoising Diffusion Probabilistic Model (DDPM) and scrutinizing the posterior distribution of the observed images, acting as a fidelity term to ensure content consistency during the iterative process. The results manifest the remarkable performance of the CPDM despeckling and confirm the effectiveness of CPDM for future clinical analysis. The study is cogently written with an in-depth literature review, demonstrating a considerable enhancement in despeckle performance on the AS-OCT. The authors affirm the preeminence of CPDM for speckle removal through a successful ciliary muscle (CM) segmentation and scleral spur (SS) localization task. Nonetheless, the paper exhibits a few areas of concern. There are questions raised about its claim of being unsupervised, given the need for clean images x_0 in DDPM training, which deviates from a typical ‘unsupervised’ framework. The paper could further benefit from providing clarification about the original image resolution size prior to resizing and a more comprehensive explanation of speckle noise as multiplicative noise. Additionally, several references are lacking key details, such as page and volume numbers. In their rebuttal to the reviewers’ comments, the authors show appreciation for the insightful feedback and offer a comprehensive response. They defend their noise model with references to previous studies, indicating a strong correlation between speckle noise and the signal in the OCT system. They commit to improving the speckle noise model’s description by incorporating the log-transform distribution of speckles. The authors further differentiate their CPDM from the existing DDPM, underlining its unique design for unsupervised image despeckling that preserves content. They clarify their process of adapting speckles into additive noise via a logarithmic transformation to accommodate the diffusion process, and the integration of a fidelity term derived from the posterior distribution of speckled images to achieve content consistency. Other points addressed include the employment of unpaired data learning as a form of unsupervised learning, the clinical benefits of preserved content in AS-OCT, and their rationale for the selective display of results due to page constraints. Unfortunately, they are unable to share the dataset and code due to confidentiality requirements, yet they express a willingness to engage in further discussions. In conclusion, given the novel and promising CPDM despeckling algorithm, the thorough and constructive responses to the reviewers’ comments, and the evident efforts made to address the concerns raised, I recommend accepting this paper.