List of Papers Browse by Subject Areas Author List
Abstract
Due to various physical degradation factors and limited photon counts detected, obtaining high-quality images from low-dose Positron emission tomography (PET) scans is challenging. The Denoising Diffusion Probabilistic Model (DDPM), an advanced distribution learning-based generative model, has shown promising performance across various computer-vision tasks. However, currently DDPM is mainly investigated in 2D mode, which has limitations for PET image denoising, as PET is usually acquired, reconstructed, and analyzed in 3D mode. In this work, we proposed a 3D DDPM method for PET image denoising, which employed a 3D convolutional network to train the score function, enabling the network to learn 3D distribution. The total-body 18F-FDG PET datasets acquired from the Siemens Biograph Vision Quadra scanner (axial field of view > 1m) were employed to evaluate the 3D DDPM method, as these total-body datasets needed 3D operations the most to leverage the rich information from different axial slices. All models were trained on 1/20 low-dose images and then evaluated on 1/4, 1/20, and 1/50 low-dose images, respectively. Experimental results indicated that 3D DDPM significantly outperformed 2D DDPM and 3D UNet in qualitative and quantitative assessments, capable of recovering finer structures and more accurate edge contours from low-quality PET images. Moreover, 3D DDPM revealed greater robustness when there were noise level mismatches between training and testing data. Finally, comparing 3D DDPM with 2D DDPM in terms of uncertainty revealed 3D DDPM’s higher confidence in reproducibility.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/3664_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: N/A
Link to the Code Repository
N/A
Link to the Dataset(s)
https://ultra-low-dose-pet.grand-challenge.org/Dataset/
BibTex
@InProceedings{Yu_PET_MICCAI2024,
author = { Yu, Boxiao and Ozdemir, Savas and Dong, Yafei and Shao, Wei and Shi, Kuangyu and Gong, Kuang},
title = { { PET Image Denoising Based on 3D Denoising Diffusion Probabilistic Model: Evaluations on Total-Body Datasets } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15007},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose a 3d DDPM for PET denoising and show that it outperforms 2d DDPM and 3d Unet in PSNR and SSIM. Authors also show 3d produces lower uncertainty in a single subject than 2d DDPM.
- 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.
- Article is well-written.
- Authors propose an approach that could be useful on many tasks. Besides the time/memory constraint, 3d networks are the intuitive choice for 3d problems like medical images. Since diffusion models have gained popularity, 3d networks within DDPM could be very useful.
- The authors evaluate the proposed method on alternative low-dose input to determine robustness.
- The authors use a decent size for test set and report statistical tests for significance on metrics.
- 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 compares against Unet as baseline after mentioning known limitation of blurring in non-GAN network.
- 3D GAN network would have been a better comparison point.
- Authors don’t mention availability of code and some training details are lacking: which “mixed precision training techniques” were used? Which optimization algorithm was used? Which loss was used?
- Evaluation metrics are generic image metrics without any clinical metrics. “Qualitative assessment” appears limited. Were other images visually graded or compared in test set or was only this single image looked at?
- Novelty is limited since this is not the first use of 3d networks within DDPM for medical images: Z. Dorjsembe, et al. “Three-Dimensional Medical Image Synthesis with Denoising Diffusion Probabilistic Models”. Medical Images in Deep Learning 2022. However, this prior work does not compare against 2d DDPM and doesn’t use it for a conditional image generation task.
- 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 provide sufficient information for 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
An error or relative error map would have been useful to visualize the error in the figures and to help determine if uncertainty map reflects the error. If space is a limiting factor, Fig. 4 A and C could be combined. Since there is only one not-significant difference, notations of ***, ns could be removed to un-crowd figures, and this could be added as a single sentence instead. Alternatively, Fig. 4D could be moved to a sentence description. Arrows on coronal or axial view in Fig. 3 would help reader find other areas of significant change. In its current state, the additional views have limited value since commentary is not detailed on differences.
- 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?
Methodology could be valuable for many applications, and this shows a proof of concept about PET denoising.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
This proposed extended the 2D DDPM into 3D version for PET image denoising, the conditional DDPM was used here with the input of noisy PET image. The proposed model was only trained on 1/20 low-dose imaging, but validated on total-body PET (1.06m AFOV, Siemens Biograph Vision Quadra) dataset with different noisy levels. Compared to 3D U-Net and 2D DDPM, the proposed showed a significant improvement, especially for edge recovery.
- 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.
- Clear description for method: As the DDPM is a well-known model, the authors give the simple but clear description to make the proposed method easily reproducibility.
- Experiment design and results: I am very impressived by the result part, like Fig. 3. In addition to the statistical evaluation across the methods. The image derived from 3D DDPM showed singnicifincant improvement than other. Particularly for the vessel wall pointed by authors too. The vessel wall is really an important factor for many disease in clinic. Other compared methods can not show its recovery. So here is the reason why I say this improvement is significant!
- Good generalization ability: Here the definition of generalization is for different noisy levels. The model was only trained on 1/20 low-dose, but can also work well for 1/50 low-dose imaging. This is another achievement which bring the field to think about the relationship between the model ability and noisy level. It’s so interesting to know this based on the author’s demonstration.
- Uncertainty map: The improvement regrading the uncertainty map from 3D DDPM is a good evidence why this method performs well and is robust in the future.
- 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 pure novelty: One would argue this work maybe a nature extendation from the previous Gong’s work (ref [8]), that is from 2D DDPM to 3D. This is true. If from the method novelty angle, it’s limited.
- Lack of other compared methods: Personally speaking, 3D-Unet comparison looks like an ablation experiment given the 3D-Unet is the backbone in DDPM. So I would say 3D U-Net comparison here is not fair to me.
- Lack of description of uncertainty map: this is the important demonstration why 3D DDPM is more trustable, the author ingorns the description of uncertainty map to make people understand what we can gain in this comparison.
- 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?
It’s clear for reproducibiltiy. As for the huge computing resources, if authors can open their trained model in the future, it would be fantastic for this field. Looking forward!
- 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
For the current draft, I have the following comments:
- Could you please give a little bit more about uncertainty map? It would be helpful to make people convincing the proposed method and then use it.
- 3D U-Net is not a good compared method as I have explained above. It would be better to include another (at least) gererative model for comparison, like GAN or others.
- As the limited the novelty of the method section, do you have some solid example to show the benefits over the proposed 3D DDPM, like for confirmed lesion recovery. It’s more important to highlight your work from the application angle.
For the future, I would recommend:
- Open the trained model if it’s available for you. This is so important beyond any method novelty for this field.
- Try to extend the generalization across different scanners, different tracers. It would be valuable to do 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?
Although the method is limited novelty, I would prefer the significant improvement from the 3D DDPM and this is a promising model to have a good generalization across different vendors, tracers and noisy levels.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #3
- Please describe the contribution of the paper
The paper describes a denoising method for PET images based on a 3D diffusion model. The method is compared with an equivalent 2D method and with 3D UNet. The proposed method presents higher values of PSNR ad SSIM. The authors demonstrated that the application of this method may lead to a significant dose reduction.
- 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 presents a very promising method for denoising PET images that may lead to a significant dose reduction, which is crucial in pediatric examinations. The method is very well explained, the differences to other known methods are well described, the methodology is adequate and the results are clear and promising.
- 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 paper does not have important weakness. However, it is not presented any metric that evaluate if there is an oversmooth of the images. This is evaluated only by visual analysis. The authors have also compared the proposed method with adaptive filters, such as Total Variation Minimization, that are commonly used for this purpose.
- Please rate the clarity and organization of this paper
Excellent
- 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 should declare if they will give access to the code used for this work
- 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 paper is already very good as it is, but it could be improved if the authors could include a quantitative analysis to evaluate any possible oversmooth of the images. The comparison with other methods could also be extended to adaptive filters, such as Total Variation Minimization, to demonstrate the benefit of training a CNN in comparison with more direct methods
- 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 paper is very well presented and structured. The presented method could lead to a significant dose reduction (increasing the potential clinical applicability). The methodology is adequate and the results are clear and well explained.
- 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
N/A
- [Post rebuttal] Please justify your decision
N/A
Author Feedback
We sincerely thank all the reviewers for providing insightful comments. We extracted the major concerns and consolidated the opinions from different reviewers. Below, we provided our responses to each point.
Reviewer #1’s fifth concern and Reviewer #4’s first concern: Both reviewers mentioned the limited novelty of our work. Reviewer 1 noted that “this is not the first use of 3d networks within DDPM for medical images: Z. Dorjsembe, et al. “Three-Dimensional Medical Image Synthesis with Denoising Diffusion Probabilistic Models”.” Reviewer 4 mentioned that “this work maybe a nature extendation from the previous Gong’s work (ref [8]), that is from 2D DDPM to 3D.” Response: Dorjsembe et al.’s work utilizes an unconditional 3D DDPM for brain MRI synthesis, while our work adapts the 3D DDPM to conditional image generation. The low-dose PET images were incorporated as additional network inputs of the score function in the training stage and guided the generation of corresponding high-quality PET images in the sampling stage. This conditioning process ensured that the DDPM’s outputs were specifically tailored for particular applications with clinical significance. Regarding Reviewer 4’s comment, Gong’s previous work only focused on 2D brain images; our research targets 3D whole-body PET images, which have a more complex data distribution due to the varying levels of radiotracer uptake across different organs, making the denoising task more demanding.
Reviewer #1’s first and second concerns and Reviewer #4’s second concern: Both reviewers mentioned the limitations of using 3D UNet as a comparison model and suggested including 3D GAN as an additional comparison. Response: We will incorporate the 3D GAN comparison results in the camera-ready file. As for the 3D UNet model, given that it served as the backbone of our approach, comparing its performance with our method is particularly insightful, which helps illustrate the advantages of the diffusion model over traditional convolutional network-based methods.
Reviewer #1’s second concern: Reviewer 1 mentioned: “Qualitative assessment appears limited. Were other images visually graded or compared in the test set, or was only this single image looked at?” Response: We conducted the qualitative evaluation on all test set data, and our method consistently demonstrated superior performance. Due to space constraints, we only presented the results of a randomly selected test image in the manuscript.
Reviewer #3’s first concern: Reviewer 3 mentioned that the paper does not present any metric to evaluate whether the images are overly smoothed. This is evaluated only by visual analysis. Response: We used PSNR and SSIM for quantitative evaluation. SSIM considers luminance, contrast, and structure to assess the structural similarity between the denoised image and the ground truth image. If the denoising process overly smooths the image, it will reduce local contrast and edges, leading to a decrease in SSIM, which indicates potential oversmoothing.
Reviewer #3’s second concern: Reviewer 3 suggested, “The comparison with other methods could also be extended to adaptive filters, such as Total Variation Minimization, to demonstrate the benefit of training a CNN in comparison with more direct methods.” Response: For PET imaging, the image noise is very different from MR and CT imaging due to the limited photon count received. As a result, total variation (TV) minimization-based methods are not widely used in PET imaging, due to the patchy and cartoon shapes in the final images.
Additionally, we appreciate all the reviewers’ constructive suggestions, such as including more training details, descriptions of the uncertainty map, and relative error maps. We will address the reviewers’ concerns as much as possible within the camera-ready file’s length constraints. We are also planning to open-source our code and trained models soon.
Meta-Review
Meta-review not available, early accepted paper.