List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Sutanu Bera, Prabir Kumar Biswas
Abstract
The deep convolutional neural network has been extensively studied for medical images denoising, specifically for low dose CT(LDCT) denoising. However, most of them disregard that medical images have a large dynamic range. After normalizing the input image, the difference between two nearby HU levels becomes minimal; furthermore, after multiplying it with the floating-point weight vector, the feature response becomes insensitive to small changes in the input images. As a consequence, the denoised image becomes visually smooth. With this observation, we propose to use HU level slicing for improving the performance of the vanilla convolutional network. In our method, we first use different CT windows to slice the input image into a separate HU range. Then different CNN network is used to process each generated input slice separately. Finally, a feature fusion module combines the feature learned by each network and produces the denoised image. Extensive experiments with different state of the art methods in different training settings (both supervised and unsupervised) in three benchmark low dose CT databases validates HU level slicing can significantly improve the denoising performance of the existing methods.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16446-0_56
SharedIt: https://rdcu.be/cVRTZ
Link to the code repository
N/A
Link to the dataset(s)
N/A
Reviews
Review #1
- Please describe the contribution of the paper
The authors improve the performance of denoising CNN in low-dose CT by slicing the problem into different dynamic range levels, significantly improving the performance of different denoising networks.
- 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.
-
This is a very good idea that solves a complex problem with a simple add-on to denoising strategies.
-
The results are very good and have been validated with different networks and different 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.
-
The threshold selection and number of dynamic of the ranges seems ad-hoc
-
Slicing the problem adds complexity to the networks that may become more difficult to training and with more parameters and hyperparameters
-
- 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
ok
- 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/2022/en/REVIEWER-GUIDELINES.html
In fig. 5 one subfigure seems to be missing Table 3 could have in bold the best results for a better visualization A discussion on how to choose number of ranges and the thresholds would be interesting
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is an idea that can be widely applied for many different neural networks and approaches, even for different applications. The enhancement of performance of the is very significant.
- Number of papers in your stack
1
- What is the ranking of this paper in your review stack?
1
- 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
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #2
- Please describe the contribution of the paper
Authors proposed the use of HU level slicing to improve the performance of the vanilla convolutional network. Authors first use different CT windows to slice the input image into separate HU range. Then different CNN networks are used to process each slice. Then, a feature fusion module combines the feature learned by each network and produces the denoised image.
Quantitative (PSNR, SSIM, RMSE) and qualitative experiments are presented comparing the proposed technique and state-of-the-art methods. The experiments show an good performance (PSNR, SSIM and RSME)
- 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: The novelty of the work lies in the use of HU level slicing on CNN networks.
- The method shows a good performance (in terms of PSNR, SSIM and RSMEI) when combined with some state-of-the-art methods.
- 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.
- Some important information is missing in the experiments: The computational time is not presented. A declaration of what software framework and version authors used is not presented. A description of the computing infrastructure used (hardware and software) is not presented.
- English could be improved.
- 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 average runtime for each result, is not presented.
- A declaration of what software framework and version authors used is not presented.
- A description of the computing infrastructure used (hardware and software) is not presented.
- 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/2022/en/REVIEWER-GUIDELINES.html
- It would be interesting that authors gave some information about the computational time required by the proposed method.
- Authors should include a description of what software framework and version used.
- Authors should include a description of the computing infrastructure used.
- Give a reference (and/or formulas) for RMSE, PSNR, SSIM.
- Summarize the research limitations and future research directions.
- English needs a revision.
- 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
6
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- The topic of the paper is relevant and interesting to the MICCAI community.
- It presents an innovative idea (HU level slicing ) that can be used in existing methods to enhance the denoising performance in Low Dose CT Denoising.
-
Some important information is missing in the experiments, but it’s not complicated or expensive for authors to include it. This missing information affects to the reproducibility of the experiments.
- The code is not available, but enough details to reproduce it are given.
- Number of papers in your stack
6
- What is the ranking of this paper in your review stack?
5
- 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
6
- [Post rebuttal] Please justify your decision
Authors will add the required information in the main paper. Although the idea of HU level slicing has been used before in CT image analysis, the way the authors used the HU level slicing to improve the denoising process is of sufficient interest to be accepted. The topic of the paper is relevant and interesting to the MICCAI community.
Review #3
- Please describe the contribution of the paper
The authors point out the importance of intensity resolution in LDCT. To tackle the issue by large dynamic range in LDCT and improve the deep neural network for LDCT denoising, the authors propose to use HU level slicing where feature fusion mechanism from multiple deep neural network for different HU levels is developed.
- 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 main strengths of the paper is that they bring up the intensity resolution issue in LDCT denoising. In particular, most deep neural network uses min/max normalization for preprocessing of the input image, the relatively large dynamic range needs to be adapted for the LDCT denoising network. The proposed method is validated on multiple public 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.
The main weakness of the paper is that HU level slicing is not novel method. This is the traditional thresholding with heuristic parameters. The method needs ablation study for HU slice values and theoretical analysis why this HU level slicing is beneficial. Most importantly, HU values have a physical semantics in terms of attenuation so normalizing after HU slicing needs to be justified. In particular, the reconstruction loss is well balanced between different HU bins.
- 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
Since the authors used or updated the existing network such as U-Net, REDCNN and CycleGAN, it is easy to reimplement without source code release. It will be trivial to implement HU level slicing if the HU level values are given.
- 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/2022/en/REVIEWER-GUIDELINES.html
The authors need to justify why HU value slicing is working theoretically. In addition, they need to discuss the physical meaning of HU value in terms of attenuation. Moreover, they need ablation study with different HU values and discuss how to set HU values. Theoretically, they can make infinite number of HU slices and their linear assumption on performance with the number of HU slices may be violated. They also need to consider weighting in terms in L_recon for different slices. Moreover, learning procedures and implementation details after fusing the features should be illustrated.
- 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 major factor for my recommendation is the novelty of the proposed method. HU level slicing is heuristic and validated with many different settings for robustness check. Theoretical analysis should be included for justification of using HU level slicing.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
4
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
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.
This paper proposes HU level slicing to improve the performance of convolutional network based CT image denoising. The reviewers agree that the idea is interesting and the achieved performance is promising.
-
The threshold selection and number of dynamic of the ranges seems ad-hoc. Please clarify how to choose number of ranges and the thresholds. HU values have a physical semantics in terms of attenuation so normalizing after HU slicing needs to be justified.
-
The idea of HU level slicing has been explored before in CT image analysis, for example https://doi.org/10.1109/JBHI.2019.2946066
-
Details of the experimental setting are unclear.
The authors are invited to provide a rebuttal.
-
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and 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).
4
Author Feedback
First of all, we would like to thank all the reviewers for their constructive comments and suggestions. Although we had included all the detailed training settings in the supplementary material, nevertheless, as requested by R#2 and Meta Reviewer 1, we will add all the basic information in the main paper. Regarding the number of slices: The procedure to select the range and threshold for each slice is already explained at the starting of section 2. The number of slices is an adjustable parameter; it can be modified accordingly to the design choice. If k number of slices is used, then the width of each window will be (4096/k), and the range of the starting window will be [-1024, (-1024+4096/k)]. It is intuitive that with more slices, the intensity resolution of each slice will increase; consequently, the problem of low-intensity resolution will mitigate. We found that, with five-level slicing, the performance gain is quite noteworthy for LDCT denoising. In Table 2, we have shown a sensitivity analysis of how the result changes with a change in the number of the slicing window. Accordingly, in all the experiments of section 4.2, we have used five slices(we have also mentioned it in section 4.2); thus, the width of each window is ~800. Therefore, the range of the first window is [-1024, -400]; similarly, the range for the remaining windows can be calculated. For other tasks, different number of slices can be used, and accordingly range of the width of each window can be calculated. R#3 raised a point regarding the justification of normalization after HU slicing. We would like to state that normalization is a standard step used before feeding the input data to a neural network. Our input data is slices of the input image, so we normalized the slices. Regardless, we preserved the relative contrast of HU bins throughout the network and separately predicted HU slices of each window. The predicted HU slices were denormalized and then combined to produce the final denoised image. Next, we admit that none of the HU slices has complete semantic knowledge about the input image, so we combined the latent space features using a novel masking-based feature fusion module. Now, the reconstruction loss can be modified by giving different weights to different slices; however, we wanted to keep the method as generic as possible, giving equal weightage to each HU bin. A more exclusive choice of weights will produce a more satisfactory result. Another point raised by R#3 is the novelty of the HU level slicing. We agree that CT windowing and slicing are not new to the medical image community. However, how we have used the HU level slicing to improve the denoising performance of the existing SOTA method is entirely novel. To the best of our knowledge, none of the image denoising/restoration methods has investigated this aspect. We are also the first to benchmark the effect of the low-intensity resolution or extensive dynamic range of medical images in the neural network. In section 4.1 (sensitivity analysis), we have discussed it in detail. To mitigate the low-intensity resolution problem of the medical image, we have proposed to employ HU level slicing. Our proposed method is straightforward and generic and can be used to extend any method irrespective of whether it is a supervised adversarial/MSE loss based or unsupervised method. We have demonstrated the performance gained by our method is very promising, and it will create a new direction for medical imaging restoration. Meta-Reviewer 1 provided the earlier literature on medical image classification, which also used an approach to segment the input image into different HU bins. Apart from that, our method is thoroughly different from theirs. After windowing, they concatenated the HU bins and used them as input. Concatenating the HU bins may disturb the relative difference between different HU values (the range of each bin is [0,1]); as a result, it might fail to achieve efficient denoising.
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 paper is of good quality as agreed by the reviewers. The rebuttal further clarifies some of the concerns raised in the previous round of review, e.g. novelty. The paper is thus recommended to accept.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
2
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.
This paper presents a generic, novel, and highly effective HU level slicing method to improve SOTA CNN approaches for low-dose CT denoising. The majority of the reviews are favorable, and the authors were able to address the concerns raised by Reviewer 3 on threshold selection and novelty of the paper method. I believe this paper is highly relevant to MICCAI and a novel contribution to low-dose CT denoising.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- 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).
1
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.
The rebuttal has succesfully clarified the concerns from the reviewers. The proposed method has been validated on multiple datasets with good performance.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Reject
- 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).
3