List of Papers Browse by Subject Areas Author List
Abstract
Deep image prior (DIP) has become an important approach to unsupervised reconstruction of Positron Emission Tomography (PET) images. In the setting of dynamic PET, however, its performance is limited by the frame-by-frame reconstruction, computational cost, and the fixed-size discrete representation of PET images. To address these challenges, we propose IMREPET, a novel dynamic PET reconstruction method based on implicit neural networks (INR). By incorporating temporal information directly into INR’s parameterization of dynamic PET images, we overcome the limitation of frame-by-frame reconstructions without the need of complex algorithms or regularization. Results on simulated and real-data experiments showed that IMREPET enabled rapid, high-quality reconstruction with improved signal-to-noise ratio and enhanced image detail recovery, while drastically reducing computation time compared to DIP baselines. The resolution-agnostic nature of INR further allowed IMREPET to reconstruct PET images at any resolution. These results show the feasibility of IMREPET as a robust and efficient solution for dynamic PET imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2327_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)
Brainweb dataset: https://brainweb.bic.mni.mcgill.ca/anatomic_normal_20.html
BibTex
@InProceedings{FanKai_IMREPET_MICCAI2025,
author = { Fan, Kailong and Ye, Yubo and Liu, Huafeng and Wang, Linwei},
title = { { IMREPET: Implicit Neural Representation for Unsupervised Dynamic PET Reconstruction } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15961},
month = {September},
page = {269 -- 279}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes IMREPET, the first framework leveraging Implicit Neural Representations (INR) for dynamic PET reconstruction. IMREPET incorporates unsupervised learning to provide a strong prior for image reconstruction, significantly reducing computation time compared to DIP-based approaches. Additionally, the method supports reconstruction at arbitrary resolutions, including the capability to achieve super-resolution during the reconstruction process.
- 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 outperforms existing approaches across all count levels and evaluation metrics. It demonstrates significantly faster computation times compared to DIP-based methods. Additionally, the method is validated on both simulated and real datasets, highlighting its practical applicability.
- 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 baseline methods used for comparison are not up to date, limiting the strength of the experimental evaluation. Moreover, the use of only a single real dataset for both training and testing raises concerns about the generalizability of the proposed approach.
- 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
There is only one subject in the real dataset, which raises concerns about how the training and testing sets are separated. I recommend evaluating the method on additional subjects to better assess its generalizability. Furthermore, for real patient data, it is unclear how scatter and random events are handled—this should be clarified.
In the Introduction, the authors mention state-of-the-art kernel methods that incorporate deep learning (14, 15). However, in the experimental comparison, only the original kernel method is used as a baseline, which does not reflect the most recent advancements in the field.
According to Reference (6), the evaluation of DIP-based reconstruction methods appears more promising than what is reported for DIPNMF in this paper. In Figure 4, the DIPNMF method shows a suboptimal tradeoff between standard deviation and bias. The authors should explain the cause of this performance degradation.
In Figure 6, the DIPNMF result for the Thorax case appears to produce the worst visual quality and shows structural discrepancies compared to other methods. However, in the quantitative evaluation using the simulation dataset, DIPNMF achieves results that are comparable to or even better than KEM and MLEM. The authors should address this discrepancy between simulation and real human findings.
There is a typo in the Introduction on page 1: “te nd” should be corrected to “tend.”
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well-written and clearly organized; however, more recent state-of-the-art methods should be included in the comparison. Additionally, the performance of the proposed NMF-based approach should be evaluated on a larger set of real patient data to better assess its clinical applicability.
- 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 #2
- Please describe the contribution of the paper
This paper proposes an INR-based framework for dynamic PET reconstruction. This framework is an unsupervised paradigm that does not rely on any prior images and maintains efficiency compared to existing DIP-based methods. Quantitative experiments on both simulated and real-world datasets demonstrate the effectiveness of the proposed framework.
- 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 major strengths are:
-
An unsupervised framework that utilizes the INR for dynamic PET reconstruction.
-
The proposed framework does not require any prior images and maintains efficiency compared to existing DIP-based methods.
-
Benefiting from the continuous nature of INR, the proposed framework can achieve any resolution reconstruction, including super-resolution.
-
Quantitative experiments on both simulated and real-world datasets are conducted to evaluate the effectiveness of the proposed framework.
-
- 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 main weaknesses are:
-
There is a lack of exploration and discussion of alternatives to the key components of the proposed framework. The neural network is based on MLP with ReLU activations followed by a Sigmoid activation and incorporates Positional Encoding by Fourier Encoding. Such an architecture is considered a baseline in the INR reconstruction task [1, 2], which is too fundamental. Meanwhile, the ablation studies in Fig.7 are mainly about the (a) incorporate/ablate t, (b) convergence iterations on different G values and (c) PSNR box-plot with different G values. However, the activation function of MLP and extra regularization are not well explored and considered. For example, the SIREN [1] (with Sin activation) or WIRE [2] (with Wavelet activation) generally outperforms the ReLU baseline. Will the framework be further improved by incorporating these different activation functions in INR? Besides, the regularization, such as Total Variation (TV), also plays a crucial role in the INR-based reconstruction method. Incorporate a regularization could further improve performance and convergence.
-
There is a typo in Eq.(2), where the “!” is inappropriate at the end.
-
The abbreviation “PET” does not have a full name in the first place it appears (Abstract or Introduction). I guess it is “Positron Emission Tomography”. Its full name should be included appropriately for better clarity.
References
[1] V. Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein, ‘Implicit neural representations with periodic activation functions’, in NeurIPS, 2020. [2] V. Saragadam, D. LeJeune, J. Tan, G. Balakrishnan, A. Veeraraghavan, and R. G. Baraniuk, ‘WIRE: Wavelet implicit neural representations’, in CVPR, 2023, pp. 18507–18516.
-
- 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.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please refer to major strengths and weakness.
- 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
This paper proposes a novel approach to unsupervised dynamic PET reconstruction based on Implicit Neural Representations. The method trains an INR directly on the sinograms, using a Poisson log-likelihood loss, and outputs activity information as function of the space and time.
- 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 method demonstrates strong performance, both qualitatively and quantitatively, while maintaining a reasonable computation time (~100 seconds), which is far better than existing deep learning methods such as DIPNMF.
The paper is clearly written and easy to follow. All essential implementation and experimental details are provided.
The small-scale ablation study (on the inclusion of t as input, the number of iterations, and Fourier encoding dimensions) is appreciated and contributes to the overall understanding of the method.
The inclusion of real patient data is also appreciated.
- 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 experimental evaluation is limited to two simulations and one real patient scan, which is preliminary. For a dynamic PET experiment time activity curves should also be shown.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
I have a question regarding the training stability when using the Poisson log-likelihood as the loss function. This objective is known to exhibit instability near zero due to its non-Lipschitz gradient. Although the authors mention the use of gradient clipping and an aggressive learning rate decay schedule, I would be interested to hear more about how these issues manifested during training. Was the loss indeed unstable at times?
Another question concerns the super-resolution aspect. While the authors tested spatial super-resolution, I wonder whether temporal super-resolution could also be achieved — perhaps by treating time as an explicit input to the INR like the authors have tested, even if it slows down the training. Again for the super-resolution study, I would suggest to diminish the resolution in the sinogram rather in the image; the image represents the patient, not discrete, the sinogram is discretised because of the limited resolution.
Additionally, there appears to be a minor typo in Section 2 (“Optimization”): “Given observed sonograms” -> sinograms In equation (6), consider usind either “truth” or “true” but not both.
While the results remain preliminary due to the limited dataset, they are nonetheless promising. For future work, I would encourage the authors to conduct a more extensive evaluation on a larger and more diverse database, along with additional ablations — for example, assessing sensitivity to the number of projections.
- 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?
This is a well-written and well-executed paper with impressive preliminary results. The authors propose a comprehensive evaluation of their method, with several interesting ablations, and the presentation is very clear. While the application of INRs to PET imaging is not entirely novel, extending this concept to dynamic PET reconstruction is a logical and interesting step forward. The main limitation lies in the limited experimental validation.
- 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
Author Feedback
We would like to express our sincere appreciation to all three reviewers for their thoughtful and constructive feedback. We have carefully reviewed and categorized the major suggestions raised, and we provide detailed responses to each point as follows.
- The training and testing sets splitting for real data: Regarding Reviewer #1’s concern about the difficulty of separating training and testing sets when using only one real subject, we would like to clarify that our method is learning-free and does not require separate training. The reconstruction is performed directly on the real sample without any prior training phase, so there is no issue of data splitting between training and testing sets.
- Further explanation on the degradation of DIPNMF performance: Reviewer #1 pointed out that although the DIPNMF method performs well on global image quality metrics (PSNR and SSIM), it shows relatively poor performance on ROI-level metrics (trade-off between SD and bias). Since the original DIPNMF paper did not report ROI-level metrics, there is no standard reference for comparison. We believe this limitation may be attributed to the complexity and limited generalizability of the DIPNMF model. Specifically, DIPNMF combines multiple U-Nets with an NMF-based framework, leading to many hyperparameters that require manual tuning (e.g., the number of U-Nets and the weighting of different loss terms). To ensure a fair comparison, we used the same parameter settings across all simulated subjects, which may have introduced suboptimal trade-offs when averaged across diverse samples. This issue may also explain why DIPNMF performed less favorably on real thorax data. For fairness, we used the same parameters for both synthetic and real data across all methods. Given that DIPNMF uses different U-Nets to represent regions with different tissue characteristics, more U-Nets may be required to better capture the complexity of real thorax data, which could potentially improve its performance.
- For scatter and random events correction on real data: For images reconstructed using the manufacturer’s standard reconstruction technique, the correction follows the methods used as the clinical standard, which is the same as the reference[21] in our paper.
- Some suggestions for our work: Both Reviewer #1 and Reviewer #3 suggested that we should validate our method on additional datasets. Reviewer #1 suggested replacing the original kernel method with more recent kernel approaches, such as neural kernel. Reviewer #2 suggested including ablation studies about activation functions and regularization terms. Reviewer #3 suggested including Time-Activity Curves, temporal super-resolution, and reducing the sinogram resolution. We sincerely appreciate these insightful recommendations and plan to include more experiments in our future work.
- On the further discussion of training instability: We appreciate Reviewer 3’s discussion regarding the training stability when using the Poisson log-likelihood loss function. As mentioned in our manuscript, applying gradient clipping and an aggressive learning rate decay schedule significantly improves stability. Without gradient clipping, the training process is unstable, and the reconstruction performance indeed degrades. With appropriate parameter settings and the use of gradient clipping, as shown in Figure 7 of our paper, both the loss and the PSNR of the reconstructed images remain relatively stable over time. Naturally, a more detailed discussion on this topic would require further experiments, which we plan to explore in future work. We appreciate your valuable suggestion.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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