List of Papers Browse by Subject Areas Author List
Abstract
Diffusion models (DMs) have recently been introduced as a regularizing prior for PET image reconstruction, integrating DMs trained on high-quality PET images with unsupervised schemes that condition on measured data. While these approaches have potential generalization advantages due to their independence from the scanner geometry and the injected activity level, they forgo the opportunity to explicitly model the interaction between the DM prior and noisy measurement data, potentially limiting reconstruction accuracy. To address this, we propose a supervised DM-based algorithm for PET reconstruction. Our method enforces the non-negativity of PET’s Poisson likelihood model and accommodates the wide intensity range of PET images. Through experiments on realistic brain PET phantoms, we demonstrate that our approach outperforms or matches state-of-the-art deep learning-based methods quantitatively across a range of dose levels. We further conduct ablation studies to demonstrate the benefits of the proposed components in our model, as well as its dependence on training data, parameter count, and number of diffusion steps. Additionally, we show that our approach enables more accurate posterior sampling than unsupervised DM-based methods, suggesting improved uncertainty estimation. Finally, we extend our methodology to a practical approach for fully 3D PET and present example results from real [18F]FDG brain PET data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4453_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{WebGeo_Supervised_MICCAI2025,
author = { Webber, George and Hammers, Alexander and King, Andrew P. and Reader, Andrew J.},
title = { { Supervised Diffusion-Model-Based PET Image Reconstruction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces PET-DEFT, a supervised diffusion-model (DM) approach for PET image reconstruction. Unlike prior unsupervised DM methods, PET-DEFT explicitly learns the interaction between the DM prior and noisy measurement data, improving reconstruction accuracy and fidelity to the PET data manifold. The supervised framework enables more accurate posterior sampling than unsupervised DM approaches, as demonstrated by higher sample consistency (lower LPIPS) and better adherence to the true image manifold. Also, The method is extended to fully 3D reconstruction using a memory-efficient 2.5D conditional network, with demonstrated feasibility on real FDG brain 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.
This paper makes significant contributions to the field of PET image reconstruction through several key strengths. The most notable advancement is the development of PET-DEFT, the first supervised diffusion model framework specifically designed for PET reconstruction. Unlike previous unsupervised diffusion approaches that treat the denoising process and data consistency steps separately, this method innovatively learns their interaction through end-to-end training with paired image and sinogram data. This supervised approach enables more accurate integration of the diffusion prior with measurement data, leading to improved quantitative reconstruction quality while maintaining the benefits of generative modeling.
A particularly compelling aspect of this work is its successful application to real 3D clinical PET data. The authors have developed a practical implementation strategy that combines a full 3D system matrix with an efficient 2.5D conditional network architecture. This innovative design significantly reduces computational demands while preserving crucial 3D contextual information, making the method feasible for clinical use with standard GPU hardware.
Another major strength lies in the comprehensive evaluation framework. The authors have conducted extensive experiments using a large dataset of 900 training cases, far exceeding the scale typically used in similar studies.
- 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.
While this paper presents significant advancements in PET image reconstruction using supervised diffusion models, several limitations and areas for improvement should be noted:
The methodology section lacks detailed information about the data sources used in the study. Although the authors mention using 900 training datasets derived from 39 3D scans, they do not provide sufficient details about the origin and characteristics of these datasets. A clearer description of the data collection process, including scanner specifications, acquisition protocols, and patient demographics (for real clinical data), would strengthen the reproducibility and clinical relevance of the findings.
In the presentation of results, Figure 2’s use of irregular intervals on the x-axis for dose levels is unnecessary and potentially misleading. A standardized, linear scale would provide clearer visualization of the performance trends across different dose levels. Similarly, Figure 1’s schematic diagram of the methodology appears visually unbalanced, with color schemes and layout that could be improved for better clarity and professional presentation.
Regarding methodological innovation, the paper claims adaptation of the DEFT framework but does not sufficiently elaborate on the specific modifications made for PET reconstruction. A more detailed comparison with the original DEFT approach, highlighting precisely which components were changed or optimized for this application, would help readers better assess the true novelty of the proposed method. This is particularly important as the innovation claims appear somewhat incremental without such clarification.
The paper also lacks discussion about potential innovations in the pre-training methodology. Given that the diffusion model requires pre-training on high-quality PET images, a more thorough explanation of this process and any novel aspects would be valuable. The absence of this information makes it difficult to evaluate whether the pre-training stage contributes significantly to the overall performance improvements.
The claim of being the first supervised DM-based approach for PET reconstruction requires further justification, as many existing PET reconstruction methods employ supervised learning paradigms. The authors should more clearly differentiate their work from previous supervised approaches and explain why their use of diffusion models in a supervised framework constitutes a novel contribution rather than simply combining existing concepts.
Finally, while the experimental results are comprehensive, the comparison with other methods could be strengthened through more systematic qualitative analysis. The inclusion of detailed comparative tables summarizing key performance metrics across all evaluated methods would provide readers with a clearer understanding of the relative advantages and limitations of PET-DEFT. Currently, the results are presented primarily through figures, which may obscure some important differences between methods.
- 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 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.
(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?
This paper proposes a technically competent supervised diffusion model approach for PET reconstruction, but its overall impact is limited by moderate innovation and presentation issues. While the adaptation of DEFT framework to PET is methodologically sound, the novelty claims require stronger differentiation from existing supervised reconstruction methods. The experimental results, though demonstrating competitive performance, lack in-depth qualitative comparisons with key baselines, and the visual presentation (particularly Figures 1 and 2) could be significantly improved for better clarity. The work makes a useful incremental contribution to the field but would benefit from more rigorous validation and clearer articulation of its conceptual advances beyond prior art.
- 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.
The authors have adequately addressed most of the methodological concerns raised during the initial review, particularly in clarifying the technical innovations and model validation procedures. However, their responses regarding the dataset limitations remain superficial and insufficiently substantiated. The use of a small, single-site PET imaging dataset may conflict with the authors’ claims of broad model generalizability.
Review #2
- Please describe the contribution of the paper
This paper proposed a new diffusion model-based PET image reconstruction by considering the data consistency. The process can be generally described as below: use a pre-trained score function to get an estimate (should be image); use a rule-based normalization method (based on PET raw data) to obtain the data fidelity update; input Gaussian noise, the estimated image, and data fidelity term together to train a new score function to learn the resdual;, finally use a non-negativity module (with resdual and Gaussian noise as input) to output the final trained score that is used for reconstruction. The experiments are conducted based on the simulation and real data. The compared method includes traditional iterative methods, unsupervised methods, and a model-based deep learning methods. The results show that the proposed method outperforms the unsupervised methods and performs comparably with the model-based method.
- 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.
This method proposed a new way to consider the data consistency in PET image reconstruction by using the measurement normaliztion, and also proposed a new mechanism to enable the non-negativity output to get the reasonable results. The experiment is designed 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 method seems too complicated to me. Please note, in image reconstruction, introducing too many normalization or conditions (regularization) means the risk of introducing quantification bias. This is why MLEM is still the clinical solution. In addition to results, is there a theoretical reason behind your method that demonstrates each step is necessary?
- As far as I know, the compared supervised method was proposed 5 years ago. Recently, there is a new SGD-based PET image reconstruction has been proposed and post their code online. It is important to show the better performance of your method as compared to other sorce-function-based image reconstruction methods. Please refer to “https://www.melba-journal.org/papers/2024:001.html”
- The stability of this method should be tested or discussed. As you introduced too many other operation to make the data match with the current distribution, does it work for other dataset (with different distribution). I already observed the training different between 2D and 3D (different hyperparmeters). Quantitating and measuring stability is a key concern.
- 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 submission does not provide sufficient information for 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
Small comments for your consideration:
- The eq. 3 is hard to follow. The PET measurement m is not shown in normalizetion formula. Why using u-map as the normlaization factor? is that due to the whole tissue region (I guess)?
- There is no global index shown in the paper (maybe I miss them). But it’s hard to see big difference by following the Fig. 3. Either difference image or SSIM, RMSE results maybe very helpful.
- Only your method’s result for the real data. Could you please show one compared method?
- 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?
First of all, I acknowledge the author’s methodological achievements. My major concern regarding a weak rejection relates to the robustness and stability of your method and performance comparison with other score-function-based methods.
- Reviewer confidence
Very confident (4)
- [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 paper proposes a supervised diffusion model-based approach for PET image reconstruction that integrates measured sinogram data directly into the learning process. The authors claim the first supervised diffusion model tailored for PET reconstruction that gives superior or comparable performance to state-of-the-art methods at multiple dose levels.
- 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.
This appears to be the first application of a supervised conditioning framework for diffusion models in PET image reconstruction. Previously, unsupervised approaches were limited in their ability to model interactions between noisy measurements and learned image priors. This new PET_DEFT approach bridges the gap by incorporating measured sinogram data during training. They demonstrate strong performance across a wide-range of dose levels matching or exceeding the performance of state-of-the-art supervised and unsupervised methods. They apply the method to real-world clinical data to demonstrate clinical feasibility.
- 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 validation on real data is relatively narrow and could be improved by using a broader dataset with different anatomical regions, tracers, and scanner types. How will the method perform under out of distribution conditions such as unseen scanners, acquisition protocols, or patient populations? The paper emphasizes image quality metrics and visual comparisons but does not evaluate diagnostic performance such as lesion detectability and reader agreement. Without clinical task evaluation it is unclear if the method’s improvements translate to improved clinical decisions.
- 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 provide sufficient information for 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 paper lays out a novel method with high technical quality. While there are some suggestions for improvement, those could easily go into a follow-up paper and should not limit the current results from being shown.
- 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
Author Feedback
Reviewer 1 Thank you for your review and positive comments on our significant contributions to the field. Details of data sources: full details are given in ref. [25] (anonymized for review). Presentation / use of tables: for Fig. 2, we first tried a linear-scale x-axis but the wide dose range made the plot unreadably cluttered. We therefore adopted a non-linear (but not irregular) asinh scale which clearly shows the separation of performance between conventional approaches, image-prior deep learning and fully supervised deep learning. A large comparative table would obscure rather than illuminate such performance trends. Novelty over DEFT: the novel methodology that we have integrated into DEFT includes measurement normalization and a non-negativity module (both validated with ablation studies in Section 3.4). The efficient 3D training strategy is a further novel modification to DEFT. The conclusion summarizes these novelties. Pre-training methodology: we follow the DM pre-training protocol of ref. [13], altering only the voxel mask: we use the μ-map instead of the MLEM image. This swap changes at most a global scale factor, and in our 2D simulations the two variants are near-identical (as the PET image and μ-map are both derived from the same MR image). Figs. 3-4 show that any given improvement is not limited to global quantification performance. Supervised approaches: ours is the first physics-informed supervised DM for PET reconstruction. Introduction paragraph 3 contrasts it with non-DM-based supervised models, and paragraph 4 underlines its paradigm shift from previous physics-informed unsupervised DM-based approaches. Section 3.5 highlights the advantages of our approach (easier training, and posterior sampling) over an existing supervised framework.
Reviewer 2 Thank you for your review and positive comments highlighting the novelty and high technical quality of our paper. We agree that out-of-distribution performance and clinical task measures are critical topics for a follow-up paper.
Reviewer 3 Thank you for your review and recognition of the quality of our experimental design. 7.1 Method complexity: complexity alone does not necessarily introduce quantification bias. While PET-DEFT and FBSEMnet-adv are the most complex methods in our study, they yield the lowest test-time error (NRMSE). While MLEM remains the clinical reference, the PET reconstruction research community still has a strong interest in more sophisticated regularized reconstruction. Indeed, scanner manufacturers have taken clear steps in this direction: relative difference prior regularization is in clinical use for many GE PET scanners (Q.clear); both Siemens and GE are actively pursuing AI-based reconstruction with FastPET and Precision DL respectively. The novel methodology that we have integrated into DEFT is well-motivated within the paper - see introduction paragraph 5. 7.2 Comparison with state of the art: in fact we already do compare to the exact score-based method pointed out by the reviewer, published in MELBA. It is called PET-DDS, and already features in our results Figs. 2-3. Indeed, we go even further, by also comparing to the recent PET-LiSch score-based method. While FBSEMnet is 5 years old, we equip it with a state-of-the-art neural architecture for fair comparison (FBSEMnet-adv). 7.3 Stability is a valid concern for supervised deep-learned PET algorithms in general, but not uniquely to our approach. It is not unusual to adapt the hyperparameters of a method to the reconstruction task at hand. 10.1 We used the μ-map to calculate a stable approximation for the size of the volume. The measured data m is implicitly included as it is used to calculate x_MLEM. 10.2 Global metrics SSIM and NRMSE are indeed shown in Fig. 2. 10.3 The real data results (demonstrating fully 3D reconstruction with only a 24GB GPU) are only practical because of our training innovations with our new method. Further real-data comparison is left for future work.
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”.
N/A
- 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