Authors

Chenyu Shen, Ziyuan Yang, Yi Zhang

Abstract

Low-count positron emission tomography (PET) imaging is an effective way to reduce the radiation risk of PET at the cost of a low signal-to-noise ratio. Our study aims to denoise low-count PET images in an unsupervised mode since the mainstream methods rely on paired data, which is not always feasible in clinical practice. We adopt the diffusion probabilistic model in consideration of its strong generation ability. Our model consists of two stages. In the training stage, we learn a score function network via evidence lower bound (ELBO) optimization. In the sampling stage, the trained score function and low-count image are employed to generate the corresponding high-count image under two handcrafted conditions. One is based on restoration in latent space, and the other is based on noise insertion in latent space. Thus, our model is named the bidirectional condition diffusion probabilistic model (BC-DPM). Real patient whole-body data are utilized to evaluate our model. The experiments show that our model achieves better performance in both qualitative and quantitative respects compared to several traditional and recently proposed learning-based methods.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43907-0_26

SharedIt: https://rdcu.be/dnwcD

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #3

Please describe the contribution of the paper

The paper proposes a new unsupervised method for denoising low-count PET images using a bidirectional condition diffusion probabilistic model (BC-DPM). The BC-DPM consists of two stages: a training stage where a score function network is learned via evidence lower bound (ELBO) optimization, and a sampling stage where the trained score function and low-count image are used to generate corresponding high-count images under two handcrafted conditions. The proposed method is evaluated using real patient whole-body data and compared with several traditional and recently proposed learning-based methods, and the experiments show that the BC-DPM achieves better performance in both qualitative and quantitative respects. The paper also added some adjustments in the conditional block during the sampling stage in order to improve the performance of the model.
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 proposed method is unsupervised, which means that it does not rely on paired data and can be more easily applied in clinical practice. The use of a bidirectional condition diffusion probabilistic model is an interesting approach and could potentially have applications beyond PET image denoising. The paper properly explains how the adjustments in the conditional block was made. They are able to show a thorough derivation of the assumptions they made for the model and was able to properly summarize it in their algorithms. The bidirectional condition ensures that the generated high-count image traces back to the low-count image by using the forward and backward pass of the sampling stage. The conditions are implemented in the latent space which generates images with proper semantics instead of just estimating the high count image from low count.
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 method of denoising diffusion models are computationally intensive. Maybe the authors can include a sampling time during the sampling stage. The images look smooth compared to the GT
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

The authors are able to provide the details of training architecture and hyperparameters in the paper. it would be very easy to reproduce the paper if the authors had provided the source code and some data.
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

Explainwhy bidirectional condition is better. How does it play a role in data consistency? In the methods used for Table. 1, include traditional methods like MLEM for PET for comparison. Provide reason why transferring bidirectional condition to latent space is advantageous in the introduction or the method section Include limitations of the method
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 transferring of condition to the latent space and using both forward and backward passes of the diffusion model is worth further exploration.
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

N/A
[Post rebuttal] Please justify your decision

N/A

Review #2

Please describe the contribution of the paper

The study aims to denoise low-SNR PET images in an unsupervised mode. Unlike the supervised paired methods, the case considered by the authors is more relevant to the clinical practice.
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.

Very simple and transparent description Clinical relevance in an application where not so many works are reported Seems to be the first application of score-based diffusion models to PET imaging Efficient reduction of dimensionality to apply diffusion process in the latent space
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 training algorithms resembles basic diffusion model training to a large extent. Proposed sampling stage is different and specific to PET; however, the authors rely on radon transform formalism, which loses some information each transform. Why is it needed exactly? What are the information losses associated with the transform?

The authors transition from gaussian to poisonian noise without proper rigor. Statistics change may affect the robustness of the sampling performance.

Results in Fig. 3 present all the models that I would like to see, but the explanation of “grainy-ness” is missing. I suspect some models may have been run in an inference setting, causing appearance of the blurriness in some frames. Also, some additional high-frequency blobs could be seen which are not present in the GT images. Are the authors confident that the reconstruction didn’t lead to creation of non-anatomic artifacts?

Only one dataset is reported. No discussion on generalizability or combining with other modality (PET-CT, where Radon transform is inherent).
Please rate the clarity and organization of this paper

Excellent
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

Good, but only one dataset
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 would help to have a short survey of radiologists about the quality of the reconstructions. Or, at least, add some perception-based metric. I doubt that PSNR or SSIM are any good for these kinds of grainy grayscale recons. Please include more medical and technical background, so that the references abundantly reflect the state of art in the field.
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?

I didn’t find any other works on PET low-dose imaging with diffusion models. So, despite the weeknesses, I opt for weak accept.
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

N/A
[Post rebuttal] Please justify your decision

N/A

Review #1

Please describe the contribution of the paper

BC-DPM was proposed for denoising PET images at different count levels.
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.
1. The method proposed in this article is essentially unconditional generation + conditional correction (Eq. 10), in which the correction of Eq. 10 draws inspiration from the idea of ILVR and incorporates two conditions (Eq. 6) based on the actual situation of low-count PET images.
2. The experiments are quite comprehensive.
3. The main part of the paper is easy to follow.
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.
1. In the experiments, the low-count PET images were simulated from high-count PET images. Why didn’t the paper use these simulated low-count PET images and high-count images for supervised training? This experiment was not compared in the paper.
2. The time cost of the diffusion model was not mentioned.
3. The diffusion model should be able to generate multiple denoising results, but the paper did not explain or demonstrate this.
4. The paper did not explain the impact of different l values on the denoising results.
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

Reproducibility is satisfied.
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
1. It would be beneficial to add some experiments, such as supervised learning using simulated data and investigating the impact of different l values on denoising results.
2. Further explanations should be added for some experimental details. (See Answer 6)
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?
1. The design of the two conditions is quite interesting.
2. The experimental results are good, but the persuasiveness is somewhat lacking.
Reviewer confidence

Very 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.

This paper introduces a bidirectional condition diffusion probabilistic model (BC-DPM) for low-count PET image denoising. The design of two handcrafted conditions are novel and interesting. The paper is well structured and easy to follow. Overall, the experiments are solid and convincing, although they could benefit from further image quality assessment.

Author Feedback

N/A

back to top

PET Image Denoising with Score-Based Diffusion Probabilistic Models