List of Papers Browse by Subject Areas Author List
Abstract
Denoising of 3D Optical Coherence Tomography Angiography (OCTA) for awake brain microvasculature is challenging. An OCTA volume is scanned slice by slice, with each slice (named B-scan) derived from dynamic changes in successively acquired OCT images. A B-scan of an awake brain often suffers from complex noise and Bulk Motion Artifacts (BMA), severely degrading image quality. Also, acquiring clean B-scans for training is difficult. Fortunately, we observe that, the slice-wise imaging procedure makes the noises mostly independent across B-scans, while preserves the continuity of vessel (including capillaries) signals across B-scans. Thus inspired, we propose a novel blind-slice self-supervised learning method to denoise 3D brain OCTA volumes slice by slice. For each B-scan slice, named center B-scan, we mask it entirely black and train the network to recover the original center B-scan using its neighboring B-scans. To enhance the BMA removal performance, we adaptively select only BMA-free center B-scans for model training. We further propose two novel refinement methods: (1) a non-local block to enhance vessel continuity and (2) a weighted loss to improve vascular contrast. To the best of our knowledge, this is the first self-supervised 3D OCTA denoising method that effectively reduces both complex noise and BMA while preserving capillary signals in brain OCTA volumes.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0736_paper.pdf
SharedIt Link: https://rdcu.be/dV588
SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72120-5_56
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/0736_supp.pdf
Link to the Code Repository
https://github.com/ZhenghLi/SOAD
Link to the Dataset(s)
N/A
BibTex
@InProceedings{Li_SelfsupervisedDenoising_MICCAI2024,
author = { Li, Zhenghong and Ren, Jiaxiang and Zou, Zhilin and Garigapati, Kalyan and Du, Congwu and Pan, Yingtian and Ling, Haibin},
title = { { Self-supervised Denoising and Bulk Motion Artifact Removal of 3D Optical Coherence Tomography Angiography of Awake Brain } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15011},
month = {October},
page = {601 -- 611}
}
Reviews
Review #1
- Please describe the contribution of the paper
Authors propose SOAD, a new fully-convolutional architecture for denoising and motion removal. SOAD essentially is a V-Net with a non-local block for refining denoising. Authors validate the method on one private dataset, outperforming baseline methods from literature and simpler architectures.
- 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.
- novelty: SOAD’s main strength is that it is the first 3D self-supervised learning framework for OCTA volume 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.
- Limited novelty: the proposed network doesn’t have too much modification from V-net.
- Limited expansion in 3D: Though the input is a sub-volume of 3D image, the output is still 2D image.
- Limited comparison with 2D methods: authors compare SOAD with two 3D self-supervised methods and show the superiority. However, there is no comparison with 2D methods, which is needed to show the necessity for 3D extension.
- Limited task validation: Authors only evaluate the proposed network on one OCTA dateset, which cannot reflect the generality of the network on denoising. Evaluation on public dataset like OCTA-500 and ROSE
- 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
Need more information:
- Need comparison with 2D methods: it would be more convincing to include the comparison with 2D methods, like Zhang et al. (https://ieeexplore.ieee.org/document/10041892)
- Need reference on the statement: it would help to have a deeper insight into motion artifact statement that BMA appears only in a few slices if there are some reference papers to endorse it.
- 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?
The authors claim that the inter-B-scan discontinuity and approximate independence motivate them to model the inter-B-scan 3D continuity to separate noise and signal. Thus, the comparison with normal 2D self-supervised denoising network is necessary to show the effect of 3D model. But the comparison is missing in the paper.
- Reviewer confidence
Somewhat confident (2)
- [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
- acknowledge the novelty of the proposed structure.
- acknowledge that the proposed structure is awake brain OCTA only.
- Acknowledge the inaccessability of the data. Still need more verification/evaluation about the generality of the model
Review #2
- Please describe the contribution of the paper
The paper proposed a bulk motion artifact removal algorithm on OCTA volumes. Specifically, the proposed method used the nearby frames to estimate the region of interest. The study is evaluated on seven OCTA volumes from awake mice.
- 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 reviewer acknowledges the ablation study conducted by the authors.
The problem settings are clearly presented mathematically.
- 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.
The baseline comparison is not appropriate and unfair. – Out-of-date. Most of the studies are out-of-date, with only one study [29] in 2021. However, [29] is not suitable for baseline comparison because it is a general video denoising method and does not target artifact removal. – Almost all baselines are designed only for denoising. It is unfair to compare the artifact removal performance with baselines targeting denoising. – The authors should compare studies with artifact removal. The bulk motion artifact shares very similar patterns with the saturation artifact. The authors can compare with studies on saturation artifact removal, if it is hard to find studies on bulk motion artifact removal.
The size/diversity of the dataset is insufficient to support the evaluation process. – The study is evaluated on 7 OCTA volumes from awake mice. Also, the authors did not mention the number of animal models in the study.
The scope of the study is too narrow. – The study targets bulk motion artifact removal on OCTA data in awake brains and also requires 3D data to support the algorithm.
Although the reviewer understands that GTs are not easy to obtain, evaluating the study solely based on ROI-based evaluation metrics is less convincing.
The presentation of the paper can be improved by highlighting the importance/challenge of BMA removal and the challenges in the setting of the awake brain via figures/data.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
See above weakness.
- 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?
The baseline comparison is not appropriate and unfair.
The size/diversity of the dataset is insufficient to support the evaluation process.
The scope of the study is too narrow.
- 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 paper should compare with artifact removal paper for a fair comparison. It is unfair to compare with denoising paper on both denoising and artifact removal.
The artifact in this paper shares very similar pattern to saturaction and reflection artifacts. The authors should compare with those papers.
For comparison with recent “self-supervised volume denoising” algorithm, the reviewer find some papers such as: Nienhaus, Jonas, et al. “Live 4D-OCT denoising with self-supervised deep learning.” Scientific Reports 13.1 (2023): 5760.
Review #3
- Please describe the contribution of the paper
This paper proposes a self-supervised denoising and bulk motion artifact removal method for 3D OCTA. Additionally, to improve the effect of bulk motion artifact removal, a refinement method is designed, which utilizes non-local blocks and weighted loss to enhance vascular continuity and contrast. The motivation of this study is clear, the approach is rigorous, and the experiments are extensive. However, there are issues with the description of the methods section.
- 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 employed in the article to remove BMA is quite ingenious. It achieves the regeneration of slices affected by BMA by predicting the intermediate blank slices using adjacent slices.
- The experimental design in this paper is rigorous and comprehensive, with thorough discussions on hyperparameters. The choice of comparative methods includes traditional denoising, 3D data denoising, and 3D self-supervised denoising, offering a comprehensive analysis.
- 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.
- The description of the method with symbols in Section 2.2 of the article is likely to contain errors, resulting in poor readability.
- How are background noise and foreground speckle noise removed? The last paragraph of Section 2.2 provides assumptions, but lacks theoretical support here.
- 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?
Please provide the address of the source code repository in the abstract.
- 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 the last paragraph of section 2.2, the second sentence, I think the author meant to say that y_i’ cannot fit b_i’.
- In Section 2.3, below Equation 2, the author wrote: “We also take the raw B-scans y_i′ as reference to help the convergence because ˆy_i′ is not stable at the beginning of the training.” How can this be implemented? If the raw B-scans y_i′ are used for supervision at the beginning of training according to Equation 2, wouldn’t it result in zero?
- In Section 3.1, Implementation Detail, it is mentioned that one experimental setup involves using test samples in the self-supervised training process. However, this scenario is difficult to implement in clinical settings because there is a requirement for short inference times during application. Therefore, test samples should not be involved in the model training process.
- In section 3.1, Evaluation Metrics, the authors said “we pick the masks of small vessels”, what is the definition of the small vessels?
- In the ablation experiments, as seen from the first two rows of Table 2, the utilization of NL in the corrupted B-scans experiment did not improve the DICE value of vessel segmentation; instead, it decreased it. Could the authors please explain the reasons behind this?
- 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 Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea design of this paper is ingenious, and the experiments are comprehensive. However, there are issues with the method description, and the explanation of the experimental results is still inadequate.
- 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
I think this work is not very well done, but it is still a little enlightening to the field, so continue to give 4 points.
Author Feedback
Reviewer #3: (1) Writing. Thanks. We will fix writing issues. (2) Theoretical support. The distribution of noise in OCTA is very complex without a closed-form mathematical model. Therefore, we are not able to rigorously prove (like in [2]) the equivalence between using the center B-scan for supervision and using its clean version. (3) Other comments: (a) Typo: Yes, it should be b_i’. (b) Weight: This description is about w_i’. We use the raw image to calculate the weight term for training (the line below Eq.2). (c) Clinical setting: We follow settings in related work to show the potential of the self-supervised method. In clinical settings, we can quickly finetune a pretrained model on a new noisy volume for better performance. (d) Small vessels: Most of them are capillaries, and some are small arterioles and venules. No major vein or artery. (e) Segmentation result: In rare cases, there can be consecutive corrupted B-scans. NL aggregates the neighbor BMA information for the inference, and the BMA in the center B-scan may not be perfectly removed without WL in training.
Reviewer #4: (1) Network novelty. To our best knowledge, we are the first to apply a non-local block in CNN for 3D self-supervised denoising. (2) Expansion in 3D. It is hard to design a 3D-to-3D mapping that is effective in blinding the 2D correlation of noise in OCTA volume for self-supervised learning. This is why 3D Noise2Self fails. In contrast, our 3D-to-2D mapping utilizes both 3D information and inter-B-scan independence of noise. (3) Comparison with 2D methods. Comparing with the 3D methods is for fair comparison. It is redundant to show the results of 2D methods since the two tested 3D self-supervised methods are extensions of 2D methods [2,14]. The failures of both (compared with our SOAD) show that ignoring the 2D correlation of the noise may cause overfitting to noise. (4) Other datasets. We focus on awake brain OCTA volume denoising, but there is no such public dataset. Both OCTA-500 and ROSE are retinal datasets and not for denoising. We will clarify this in the revision. (5) Other comments: (a) New method: Thanks for the suggestion. Unfortunately, the conference policy does not allow adding new results. We will discuss it in the background introduction. (b) Reference of statement: Thanks. The basis of the statement is that BMA appears in 3.6% B-scans in the data.
Reviewer #5: (1) Baselines: (a) We did not find any self-supervised volume denoising method published in major venues in the last two years. We compared our method with [29] because it aggregates 3D information for denoising. (b) As for BMA, please note that our task is for BOTH denoising and BMA removal, and both are critical in processing awake brain OCTA. So, it is fair to make a comparison of BMA removal with other denoising methods. Besides, no related work about brain OCTA BMA removal on the B-scan level is found. (c) For saturation artifacts, studies are mainly about artifacts spanning only several A-scans (columns) in OCT, while BMA may appear in a much longer range in awake brain OCTA. It is unclear and nontrivial to apply those methods to OCTA; moreover, those studies do not work for denoising. (2) Dataset size: More data are desirable but nontrivial to collect from awake mice. Previous works [10,18] use only 6 volumes (5 training and 1 testing). We use 7 volumes captured at different dates and angles from 2 awake mice. (3) Study scope and 3D data: Note that our task is for BOTH denoising and BMA removal, and 3D information is helpful for both. First, comparing BM3D with BM4D, 3D information helps BM4D to recover small vessels. Second, [25] also uses neighbor B-scan information (MIP) for BMA removal. (4) Metrics: We follow previous work [8] using ROI-based metrics. Moreover, the results of downstream segmentation tasks can compensate for the ROI-based metrics, especially for small vessels. (5) Highlighting the importance of BMA removal: Thanks. We will do it.
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’
After a thorough evaluation of the reviews and the authors’ responses, I recommend accepting this paper. The authors have introduced a novel self-supervised denoising and bulk motion artifact removal method for 3D OCTA, which demonstrates significant potential in enhancing vascular continuity and contrast in medical imaging. While there are noted issues with the methods section’s clarity and certain experimental explanations, the innovative approach and rigorous experimental validation presented are compelling. The authors’ refinement method utilizing non-local blocks and a weighted loss significantly advances the current capabilities in this field. The paper has its weaknesses, such as the incomplete theoretical support for some assumptions. However, the detailed algorithmic description aids in ensuring reproducibility to a satisfactory extent. The rebuttal provided by the authors adequately addresses many of the concerns raised during the review process, reinforcing the paper’s strengths and clarifying misunderstandings regarding the methodology and results. Hence, I believe the merits of the work outweigh the limitations, making it a valuable contribution to MICCAI.
- 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).
After a thorough evaluation of the reviews and the authors’ responses, I recommend accepting this paper. The authors have introduced a novel self-supervised denoising and bulk motion artifact removal method for 3D OCTA, which demonstrates significant potential in enhancing vascular continuity and contrast in medical imaging. While there are noted issues with the methods section’s clarity and certain experimental explanations, the innovative approach and rigorous experimental validation presented are compelling. The authors’ refinement method utilizing non-local blocks and a weighted loss significantly advances the current capabilities in this field. The paper has its weaknesses, such as the incomplete theoretical support for some assumptions. However, the detailed algorithmic description aids in ensuring reproducibility to a satisfactory extent. The rebuttal provided by the authors adequately addresses many of the concerns raised during the review process, reinforcing the paper’s strengths and clarifying misunderstandings regarding the methodology and results. Hence, I believe the merits of the work outweigh the limitations, making it a valuable contribution to MICCAI.
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’
There are some issues regarding the validation method, and a more comprehensive comparison that includes artifact removal is also necessary. Despite that, the authors have addressed most of the issues in their rebuttal, and I hope the final paper mentions the limitations of the paper.
- 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).
There are some issues regarding the validation method, and a more comprehensive comparison that includes artifact removal is also necessary. Despite that, the authors have addressed most of the issues in their rebuttal, and I hope the final paper mentions the limitations of the paper.