List of Papers Browse by Subject Areas Author List
Abstract
Low-dose Positron Emission Tomography (PET) reduces radiation exposure but suffers from severe noise and quantitative degradation. Diffusion-based denoising models achieve strong final reconstructions, yet their reverse trajectories are typically unconstrained and not aligned with the progressive nature of PET dose formation.
We propose MAP-Diff, a multi-anchor guided diffusion framework for progressive 3D whole-body PET denoising. MAP-Diff introduces clinically observed intermediate-dose scans as trajectory anchors and enforces timestep-dependent supervision to regularize the reverse process toward dose-aligned intermediate states. Anchor timesteps are calibrated via degradation matching between simulated diffusion corruption and real multi-dose PET pairs, and a timestep-weighted anchor loss stabilizes stage-wise learning. At inference, the model requires only ultra-low-dose input while enabling progressive, dose-consistent intermediate restoration.
Experiments show consistent improvements over strong deep learning baselines on both internal (Siemens Biograph Vision Quadra) and cross-scanner (United Imaging uEXPLORER) datasets. MAP-Diff improves PSNR from 42.48 dB to 43.71 dB (+1.23 dB) and reduces NMAE from 0.115 to 0.103 compared with 3D DDPM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2041_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JinPei_MAPDiff_MICCAI2026,
author = { Jing, Peiyuan AND Cheng, Chun-Wun AND Yang, Liutao AND Zhang, Zhenxuan AND Lima, Thiago A. AND Strobel, Klaus AND Leimgruber, Antoine AND Aviles-Rivero, Angelica I. AND Yang, Guang AND Montoya-Zegarra, Javier A.},
title = { { MAP-Diff: Multi-anchor Guided Diffusion for Progressive 3D Whole-Body Low-Dose PET Denoising } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposed MAP-Diff, a multi-anchor guided diffusion framework for progressive 3D whole-body PET denoising.
- 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.
By introducing clinically observed intermediate-dose anchors and timestep-dependent trajectory supervision, MAP-Diff constrains the reverse diffusion process to follow a dose-aligned and stage-consistent restoration path, addressing the mismatch between endpointonly objectives and progressive PET dose structure.
- 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.
It is suggested to include a discussion on the generalizability of the methods.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
It is suggested to discuss the timing of training and sampling.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is very interesting and the experiments have provided strong evidence of its effectiveness.
- 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.
Interesting paper.
Review #2
- Please describe the contribution of the paper
The paper addresses the structural and objective mismatches in current diffusion-based PET denoising, where the stochastic reverse trajectory often ignores the progressive physical nature of PET image formation. To bridge the gap, the authors propose a multi-anchor guided 3D diffusion framework that constrains the denoising process using clinically observed intermediate-dose scans as trajectory anchors. The model enforces stage-wise semantic consistency via a boundary calibration technique to align diffusion timesteps with real-world dose degradation and a timestep-weighted anchor loss.
- 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.
1.The method seems reasonable as there was some evidence showing that the same sample with different doses can be found and approximated in the posterior distribution during denoising process. 2.The paper is well written and not hard to follow. 3.The results are fair, with datasets from two centers, and the psnr is high compared to the comparative methods.
- 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 paper adopts Gaussian diffusion as a tractable surrogate for inherent Poisson statistics in PET image. While the boundary calibration aligns these statistically, a Poisson-based diffusion or a specific noise-injection scheme can be included as comparison and exploration, since PET noise is count-dependent.
2.The authors should include more diffusion-based PET denoising methods as comparison like: [1] Jiang C, Pan Y, Liu M, et al. PET-diffusion: Unsupervised PET enhancement based on the latent diffusion model[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023: 3-12.[2] Han Z, Wang Y, Zhou L, et al. Contrastive diffusion model with auxiliary guidance for coarse-to-fine PET reconstruction[C]//International conference on medical image computing and computer-assisted intervention. Cham: Springer Nature Switzerland, 2023: 239-249.Also, although the model only requires the ultra-low-dose input at inference, the training phase involves supervising multiple 3D intermediate volumes. The authors should discuss the impact of this multi-anchor supervision on training memory requirements and convergence time compared to standard 3D DDPM and other comparative methods.
3.There are samples of many doses for UDPET dataset, have you tried 1/50 dose as a ultra-low-dose scenario?
4.The visualization (zoom-in area) is not very clear for identifying the visual difference.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Referring the strength and weakness: The method proposed seems reasonable; The manuscript is well written and not hard to follow. The results are fair, with datasets from two centers, and the psnr is high compared to the comparative methods. Yet, there is some problems need to be addressed, like insufficient comparison, justification of the noise scheme, and some minor concerns.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The author proposed a multi-anchor guided diffusion framework named MAP-Diff for progressive 3D whole-body PET denoising.
- 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 proposed method uses clinically guided intermediate-dose anchors and timestep-dependent trajectory supervision to constrain reverse diffusion, achieving progressive 3D whole-body low-dose PET denoising with superior visual quality and quantitative accuracy
- 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.In the diffusion model, is the noise prediction based on a U-Net architecture? 2.The authors partition the timestep axis into four supervision zones. Is this partitioning uniform? 3.In the Boundary Calibration module, how is the body mask M obtained? 4.Please specify the numerical values of the weight coefficients in the loss function (Equation 7) used in the experiments. 5.A colorbar for SUV values should be added to Figure 2 to ensure that all images are displayed on the same scale.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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?
he proposed method leverages clinically guided intermediate-dose anchors and timestep-dependent trajectory supervision to constrain the reverse diffusion process, thereby enabling progressive 3D whole-body low-dose PET denoising. The efficacy of this method is substantiated by both qualitative visual assessments and quantitative evaluations.
- 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’ original work is innovative, and the rebuttal has addressed my relevant concerns.
Author Feedback
We thank the reviewers and AC for the valuable feedback. We are encouraged that the reviewers recognized the effectiveness of the trajectory-level supervision design, particularly the timestep-dependent anchor guidance and dose-consistent denoising behavior of MAP-Diff. We address the main concerns below.
Q1 (R2, Meta-R1): Gaussian diffusion vs. Poisson PET statistics. MAP-Diff does not require the diffusion forward process to be a physically exact model of PET photon-counting noise. Instead, the Gaussian forward process serves as a stochastic parameterisation of progressive image degradation. The key requirement is preservation of an ordered degradation structure across diffusion timesteps rather than exact equivalence between Gaussian and Poisson noise. Boundary calibration (Sec. 2.3) aligns diffusion states with real intermediate-dose PET images by matching degradation signatures (NMAE and SSIM within the body mask) relative to the clean full-dose image. Thus, the calibration targets statistically comparable degradation levels rather than identical noise distributions. Crucially, MAP-Diff supervises the clean prediction hat{x}_0(x_t,t) rather than the injected noise itself. Fig. 3 and Table 2 show that this trajectory-level supervision consistently improves both intermediate and final reconstruction quality over standard 3D DDPM using the same Gaussian diffusion formulation. Therefore, the observed gains arise from the proposed anchor-guided trajectory regularisation rather than the specific corruption model.
Q2 (Meta-R1): Validity of trajectory alignment and whether it is interpolation. MAP-Diff is not an interpolation framework. The stochastic reverse diffusion process remains conditioned only on the 1/20-dose input, while anchors supervise only the predicted clean estimates during training and do not replace the reverse diffusion updates. The standard noise prediction objective (L_{noise}) also remains unchanged. If the model collapsed to deterministic interpolation, accurate noise pred across diffusion timesteps would be impossible. Instead, the anchor loss acts as a trajectory-level regulariser that guides clean predictions toward dose-consistent intermediate states while preserving stochastic generative behaviour. Table 2 further supports this interpretation: constant weighting (S8), which over-constrains all timesteps equally, significantly degrades performance compared with timestep-adaptive weighting. This indicates that MAP-Diff improves dose consistency through stage-wise trajectory regularisation rather than deterministic interpolation.
Q3 (R1, Meta-R1): Generalizability and dependence on fixed anchors. MAP-Diff demonstrates cross-scanner generalisation on the external United Imaging uEXPLORER cohort under the clinically relevant 1/20-to-full-dose setting. The selected 1/10, 1/4, and 1/2 anchors are real intermediate-dose PET images available in the multi-dose dataset. Table 2 shows that one- or two-anchor variants still outperform the no-anchor baseline, indicating that MAP-Diff is not strictly dependent on any exact anchor combination, although full coverage provides the strongest guidance. Moreover, anchors are training-only; inference requires only the 1/20-dose input. MAP-Diff shows consistent improvements across both internal and external cohorts.
Q4 (R1, R2, R3): Minor clarifications. The denoiser uses a 3D U-Net DDPM backbone. Timestep zones are calibrated rather than uniformly divided. The body mask is obtained by thresholding the non-background PET region. The loss weighting coefficient in Eq. (7) is fixed to lambda=0.2 across all experiments. MAP-Diff uses the same backbone and sampling procedure as 3D DDPM; anchors are training-only targets, so inference memory usage and runtime remain unchanged. We appreciate R2’s suggestions on additional diffusion baselines and 1/50-dose evaluation, which are complementary to the current study and do not affect the proposed trajectory-supervision formulation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. 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.
The authors presented a multi-anchor guided diffusion framework to utilize intermediate-dose scans as anchors to supervise the reverse diffusion process and create a dose-aligned denoising path. All three reviewers acknowledged the novelty of the approach and its performance gain. However, several key concerns might still warrant a rebuttal.
-
Statistical Modeling (Poisson vs. Gaussian). As pointed out by Reviewer 2, the paper models the Gaussian diffusion chain as a surrogate for PET’s inherent Poisson statistics. While this is a common approximation, it might be concerning at ultra-low dose levels (e.g., 1/20 or 1/50 dose) when the low photon count makes the Gaussian approximation significantly less accurate.
-
Validity of the Trajectory Alignment. The Boundary Calibration was introduced to statistically align the intermediate diffusion outputs with intermediate dose levels. However, it remains unclear if this statistical alignment truly captures the physical dose characteristics or if it simply performs a sophisticated interpolation. Specifically, the mapping between the early trajectory and the starting dosage 1/10 might need a clearer physical justification.
-
Generalizability and Fixed Anchors: Reviewer 1 and the AC are concerned that the denoising trajectory is overly dependent on the specific anchors (1/2, 1/4, 1/10) used during training, given the 1/20 dose as the conditioning. The cross-vendor evaluation was limited to the same 1/20 dose level, which might not sufficiently demonstrate the model’s generalization.
The authors are encouraged to clarify these concerns along with other minor points given by the reviewers during the rebuttal phase.
-
- 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.
The authors clarified that their trajectory does not model the exact physical model in low-dose imaging, but rather a stochastic parameterisation of progressive image degradation. This made the foundation of their work more solid. However, they didn’t further discuss the limitation in the generalization to different dose levels, especially when the conditioning has a different dose level rather than 1/20 dose, and for the ultra low dose 1/50 level pointed out by R2. A remaining concern of mine is whether the denoising trajectory modeling in training is limited to a specific low dose input and for all other intermediate dose levels. Considering these, I am giving a weak acceptance.
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.
All reviewers are positive about this paper. After reviewing the feedback from both the reviewers and the authors, I believe the paper meets the acceptance criteria and recommend acceptance.
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.
The authors have addressed the reviewers questions.