List of Papers Browse by Subject Areas Author List
Abstract
Reconstruction of standard-dose Positron Emission Tomography is vital for clinical diagnosis, while recent diffusion-denoising probabilistic models offer strong generative capabilities, when applied to this task, they often struggle with fine detail recovery, slow inference, and inadequate cross-slice continuity in 3D volumes. To overcome these issues, we introduce WiD-PET, which employs a wavelet transform to produce smaller wavelet-transformed inputs, and thereby reduces inference time to 10% of that required by the DDPM model. Additionally, a high-frequency enhancer is adopted for reconstructing fine and rich image details. Moreover, a spatial consistency feature extractor and spatial consistency attention are implemented to enhance cross-slice continuity in 3D PET reconstructions. Evaluations across dose levels (1/20, 1/50, and 1/100) reveal that WiD-PET consistently achieves superior reconstruction quality, detail preservation, and inference efficiency. Project page: https://github.com/SwingM/WiD.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0455_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/SwingM/WiD.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LyuQin_WiDPET_MICCAI2025,
author = { Lyu, Qingcheng and Chen, Tong and Wang, Yiran and Guo, Erjian and Zhou, Luping},
title = { { WiD-PET: PET Image Reconstruction from Low-Dose Data Using a Wavelet-Informed Diffusion Model with Fast Inference } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {692 -- 701}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a diffusion model based method on the coefficients after discrete wavelet transform to generate high-dose PET data from low-dose PET data.
- Please list the major strengths of the paper: you should highlight 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 idea of denoising low-frequency and high-frequency coefficients using different methods is interesting.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The authors should use consistent notation. Please clarify the meaning of “G” and “C.” Do they correspond to “g” and “c”?
2.In the method section, the authors need to explain why only the low-frequency component is reconstructed using DDPM. Could the low-frequency components also be enhanced using the enhancer? Alternatively, could the high-frequency components be denoised using DDPM?
-
The authors should clarify whether the generated high-dose images improve performance in clinical tasks.
-
Authors need to discuss the future work.
-
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is interesting, but I have some concerns regarding its design and clinical impact.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors’ responses have provided clear explanations to my questions. I think this paper can be accepted.
Review #2
- Please describe the contribution of the paper
The paper proposed a Wavelet-informed Diffusion Model with Fast Inference (WiD-PET).By integrating Haar discrete wavelettransformation and a high-frequency enhancer, the method can reduce inference time and improve detail recovery.
- Please list the major strengths of the paper: you should highlight 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 utilizes wavelet transformations and a high frequency enhancer (HFE) module to synthesize high quality SPET images with the enhanced details.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
I don’t have any comments.
- Please rate the clarity and organization of this paper
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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is interesting and the results also seem good enough.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I don’t have any comments.
Review #3
- Please describe the contribution of the paper
The authors have proposed a new reconstruction method for low-dose 3D PET utilizing a diffusion based deep learning model. Their key contributions are:
- Using a wavelet transforms and a high-frequency enhancer (HFE) module to ensure fine detail retention in reconstructions.
- Reducing inference time significantly compared to standard DDPM by using the much smaller wavelet transformed inputs.
- Ensures cross slice continuity in 3D volumes using the spatial consistency feature extractor (SCFE) and spatial consistency attention (SCA) modules.
- Please list the major strengths of the paper: you should highlight 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 authors use wavelet transforms is an efficient manner to solve two problems: the long inference times in standard DDPM methods, and separating the low and high frequency components of data to denoise without loosing fine details.
- The spatial consistency feature extractor (SCFE) and spatial consistency attention (SCA) modules are cleverly designed to solve the inter-slice continuity problem.
- The authors perform reconstructions at three different dose levels, which gives a thorough understanding of the effectiveness of the proposed method.
- The results are reported with 3 different error metrics along with the p-values. This is good.
- The included ablation study completes off the experiment design very well.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The authors have mentioned splitting the data into test and train sets, but no mention of a validation set has been made. If the test set itself was used for hyperparameter tuning as well, this is not ideal.
- Error in Figure 2 (a), dose levels missing.
- In Figure 3, the proposed method’s output for 1/50th dose seems to worse than some of the baselines (PET-UNet) and also worse than the proposed result for 1/100 dose level. This inconsistency is hard to understand.
- The authors should consider adding error values associated with the images visualized in Figure 3 for a more complete analysis.
- Please rate the clarity and organization of this paper
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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- The authors have identified an important problems and provided an effective solution, with solid results to back-up their claims.
- The clear and concise writing aids the reader in following the proposed method and the experiments effectively.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
In general, the authors have submitted good justifications for the issues raised by the reviewers. Their proposed solution is effective and well presented.
Author Feedback
We thank all reviewers (Meta, R1-R3) for their feedback. R1 and R3 found our work interesting; R1 and R2 recognized its effectiveness and SOTA performance. We address all concerns below.
Meta-Q1: Baselines are not competitive We evaluated on the UDPET Challenge 2022 dataset, the most widely used public benchmark for standard-dose PET reconstruction. Consistent with most prior works, we focus on brain PET to enable broader and more consistent benchmarking. Based on a comprehensive literature review prior to submission (March 2025), several recent methods were excluded due to incompatibility with our setting: some use private datasets (e.g., DaD-PET [Xie et al., JNM’24]), others incorporate additional modalities like MRI (e.g., MFGD-PET [Lin et al., Med Phys’24]), and some target whole-body PET only without releasing code (e.g., DDPM-PET [Yu et al., MICCAI’24]), making reproducible comparison infeasible. While many diffusion-based PET methods (e.g., DaD-PET, MEGF-PET, GMBF-3D-PET) still rely on standard 2D/3D DDPMs, we compared with 2D/3D DDPMs that were carefully fine-tuned for PET reconstruction to ensure a fair comparison. Our evaluation also includes strong recent baselines such as CDM-GAN (MICCAI’23), while excluding earlier weaker models—e.g., in [7], EA-GAN and AR-GAN achieve PSNRs of 23.93 and 24.15 at the 1/100 dose level, notably lower than our 26.68. We believe our baselines reflect the current state of reproducible, high-performing PET methods.
Meta-Q2: High standard deviation (std) and p-values We computed paired t-tests between our method and the second-best baseline at each dose level (see Sec. 3.1, rows 5–7). The consistently low p-values confirm the statistical significance of our improvements. The high std arises from inter-patient variability - well-known in PET imaging and also evident in the low-dose baseline results (Table 1). Despite this, our method consistently achieves lower std values, indicating greater robustness.
R2-Q1: Validation split and hyperparameter tuning We adopted the CDM-GAN (MICCAI’23) data split and used only standard hyperparameters (e.g., learning rate) without tuning on test data. No additional tuning was needed, as our loss terms are naturally balanced without extra weighting. R2-Q2: X-axis error in Fig. 2(a) Thanks and will correct it. R2-Q3: Visual inconsistency in Fig. 3 The zoomed-in region of our 1/50-dose result shows clearer textures and more complete anatomical structures than both the 1/100-dose result and the Unet baseline. The corresponding PSNR/SSIM scores of this example support this: Ours (1/50): 27.64 / 0.91; Ours (1/100): 26.41 / 0.89; Unet (1/50): 24.69 / 0.84. R2-Q4: Add error values to visualized results Please see also R2-Q3 for the error values. R3-Q1: Notation inconsistency “G”/“C” denote sets of ground-truth and conditional components; “g”/“c” are their individual instances. We will clearly state this distinction in revision.
R3-Q2: Why use DDPM for LF only As mentioned in Sec. 1, DDPMs denoise uniformly in image space and struggle with HF textures and edges, especially at ultra-low doses. Fig. 2 shows our advantage over DDPM is more salient at 1/100 dose, where HF degradation is most severe. HFE is designed to fuse three HF components via attention, which does not apply to the single LF component. Thus, DDPM is best suited for LF, while HFE effectively handles HF. R3-Q3: Clinical relevance The public dataset used lacks clinical labels and does not support downstream task evaluation. Our method improves edge, structure, and fine texture recovery (Table 1) - known to potentially benefit clinical interpretation. Our validation aligns with common practice—among 50+ relevant papers published till 2024 (e.g., CDM-GAN, MICCAI’23), only two reported clinical task results, both using in-house annotated datasets. R3-Q4: Future work We will develop a unified model to handle varied dose levels within a single architecture to further improve its pracricality.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
There are several works developping diffusion models for PET, which should be compared. The current baselines are not competitive. In addition, how the p-value is computed should be discussed as the high standard deviation indicates there should be no significant difference.
- 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’
N/A
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’
N/A
Meta-review #3
- 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’
N/A