List of Papers Browse by Subject Areas Author List
Abstract
PET reconstruction is inherently ill-posed due to the extremely low signal-to-noise ratio (SNR) of raw projection data. Although supervised deep learning has achieved superior results in low-count scenarios, its clinical adoption is hindered by the reliance on high-quality paired datasets and limited generalization. While self-supervised Implicit Neural Representations (INRs) show promise for PET reconstruction, their inherent spectral bias causes MLPs to overfit high-frequency noise, yielding artifacts or over-smoothing. We propose a self-supervised framework that transforms the PET reconstruction paradigm toward explicit 3D Gaussian Splatting (3DGS). By representing the radioactivity image as a collection of controllable and anisotropic Gaussian primitives, our method effectively decouples local signal recovery from global noise propagation. We develop a differentiable pipeline incorporating the PET system matrix into Gaussian voxelization. To enhance image fidelity, we introduce a framework integrating Gaussian representation with kernel method, where a kernel matrix derived from MRI priors is incorporated into the voxelization stage. Experiments on brain and thorax datasets demonstrate that our method outperforms traditional iterative algorithms and state-of-the-art self-supervised approaches. By establishing a novel explicit representation for PET, this work provides a robust and interpretable solution for low-SNR clinical scenarios.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4528_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{ZhoZil_TOFGR_MICCAI2026,
author = { Zhou, Zilan AND Zhang, Yulin AND Wan, Yiming AND Guo, Min AND Jin, Yufei AND Lin, Jiajun AND Liu, Huafeng AND Zhu, Wentao},
title = { { TOF-GR: Unsupervised PET Reconstruction via Time-of-Flight Gaussian Representation and Kernel-Based Prior Modeling } },
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
This paper proposes an unsupervised PET reconstruction framework based on a novel 3D Gaussian representation, where the tracer distribution is modeled as a set of learnable anisotropic Gaussian primitives. The method integrates this explicit representation with a TOF-aware forward model and introduces a kernel-based anatomical prior derived from multimodal images. The work provides a new perspective on PET reconstruction by moving away from voxel/grid or implicit neural representations toward a controllable and interpretable explicit formulation.
- 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 use of 3D Gaussian representation is novel in the PET reconstruction context, and represents an interesting alternative to voxel-based or implicit neural representations. The framework is fully unsupervised, avoiding the need for paired high-quality training data, which is highly relevant for clinical deployment. The method incorporates physics-based modeling (TOF system matrix and Poisson likelihood), ensuring consistency with acquisition processes. The integration of anatomical priors via kernel methods is well-motivated and improves structural fidelity. The experimental results demonstrate consistent improvements over multiple baselines (MLEM, DIP, INR) across different noise levels, including both simulation and clinical datasets.
- 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.
Some experimental settings rely on simplified or controlled conditions, which may limit the assessment of real-world applicability. The ablation study and analysis of design choices remain relatively limited. Clarity needs improvement, reproducibility is questionable.
- 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.
- 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
1.Practical relevance and data setting: The experiments involve preprocessing steps such as downsampling of PET data, which may deviate from realistic clinical workflows. As a result, the reported performance may partially reflect an idealized setting. It would be helpful to clarify how the method performs on full-resolution data or more realistic acquisition conditions. 2.Visualization quality (Fig. 2): The current figure resolution makes it difficult to assess fine structural details and lesion boundaries. Providing higher-resolution or vector-format figures would significantly improve interpretability. 3.Dependence on initialization (k parameter): The ablation study indicates that performance is sensitive to the scaling factor k, with optimal results achieved around k=0.8.Since the initialization is derived from OSEM, this suggests that the method may rely on the quality of the initial reconstruction. In low-count scenarios where OSEM initialization is less reliable, this dependency could potentially lead to performance degradation. A discussion on robustness to initialization would strengthen the paper. 4.Ablation study could be more comprehensive: The current ablation mainly focuses on the kernel prior and a limited set of parameters. Additional analysis on key components—such as the number and distribution of Gaussian primitives, covariance parameterization, the role of TOF modeling, and the influence of TV regularization—would provide deeper insight into the method. 5.Clarity and paper organization: Some parts of the manuscript would benefit from improved clarity and organization. In addition, there appear to be minor issues in references and formatting that should be carefully checked. For example, reference 18 should be 2025 MICCAI rather than 2026.
- 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?
Overall, this paper presents a novel and interesting direction by introducing 3D Gaussian representations into PET reconstruction. While the methodological contribution is primarily in the representation rather than a fundamentally new reconstruction framework, the approach is well-motivated, and the experimental results are consistently favorable compared to existing methods. The main limitations relate to evaluation settings, limited ablation, and clarity, which can be addressed with revision. Therefore, I lean toward a weak accept.
- 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 #2
- Please describe the contribution of the paper
The paper proposed to use 3D Gaussian Splatting with a weight TV regularization term for PET reconstruction.
- 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 adopted the self-supervised approaches 3DGS for PET reconstruction. Extensive experiments on brain and thorax datasets was performed to demonstrate the effective of the proposed method.
- 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.Section 2.2 is not very clear. The proposed method used MRI to guide the reconstruction. The motivation is not clear and the calculation is very complex. 2.The setting of comparison is not given. If DIP and INR base method also use the MRI image for reconstruction. 3.The comparison of time and memory costs should be given. 4.By Table 2, the result is sensitive to hyper-parameters. However, the proposed method reconstructs each image independently. The extension of it is limited.
- 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.
- 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?
The proposed method uses MRI as prior. However, the comparison method is not aligned with the proposed method. The computational time is not given. If the method can give the results rapidly and robust to the hyper-parameter is not fully clarify.
- 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
Proposed method aims to improve spatial adaptivity and the recovery of small lesions in low‑count PET settings.
- 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 incorporation of an anatomical kernel prior follows established PET reconstruction paradigms, though its integration into a Gaussian representation is novel. Authors have conducted a rigorous comparison with the other recon methods to compare the proposed method. However, the statistical significance of the improvements in SSIM and PSNR is unclear.
- 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 framework relies on OSEM initialization for Gaussian placement and intensity scaling. Biases from the physical modelling of the imaging system will be incorporated. Another weakness in the computational time. Authors didn’t provide the comparison of the reconstruction time. Total reconstruction time, would be greater for the current method then the currently used methods, especially for the whole body reconstruction that will limit the clinical adoption. In addition, it is is unclear that how many real patient has been involved to demonstrate the methods significance using CNR.
- 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.
- 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
While the methodological contribution is sound, the work may be better positioned for a PET‑specific imaging journal. The paper would benefit from extensive comparisons with contemporary PET reconstruction times and a clearer, more critical articulation of the method’s limitations in clinical adoption.
- 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While not all components are individually novel, their integration into an explicit Gaussian, TOF-aware unsupervised PET framework constitutes a substantial contribution. The main weaknesses lie in scalability and clinical adoption of the proposed method
- 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
We sincerely thank the Area Chair and reviewers for their constructive comments and positive evaluation. Below, we briefly clarify the main concerns and outline future directions.
1.Practical relevance and clinical scalability (R1, R3): The reviewers raised concerns about the controlled setting and clinical scalability. We clarify that TOF-GR was evaluated on both BrainWeb low-count simulation data with small lesions and real clinical TOF-PET/CT data, demonstrating feasibility in quantitative and clinical scenarios. The 128×128×96 preprocessing/cropping was used to satisfy 3D memory constraints and ensure fair comparison with MLEM, DIP, and INR. Full-resolution, whole-body, and larger-scale clinical validation will be studied in future work.
2.Computational cost and scalability (R2, R3): The reviewers noted the need for computational analysis. This work focuses on establishing TOF-GR as a new explicit representation rather than fully optimizing runtime. Compared with INR’s global coordinate-to-intensity MLP, TOF-GR uses local anisotropic Gaussian primitives, reducing the burden of global implicit fitting. Our implementation runs on a single NVIDIA V100 GPU and shows lower cost than INR under the same setting. Systematic runtime and memory benchmarking will be future work.
3.OSEM initialization and hyperparameter sensitivity (R1, R2, R3): OSEM is only used to initialize Gaussian positions and intensities; the final reconstruction is optimized by the Gaussian representation, TOF-aware forward model, Poisson likelihood, TV penalty, and anatomical kernel prior. Table 2 shows that the kernel prior improves stability across different k and λ settings. Robustness to poorer OSEM initialization, especially in extremely low-count cases, will be further investigated.
4.Baseline comparison and ablation (R1, R2, R3): All baselines were evaluated under the same data setting and resolution. MLEM, DIP, and INR represent classical iterative, unsupervised CNN-based, and coordinate-based implicit reconstruction methods. The current ablation focuses on the kernel prior and key parameters, directly supporting the claims of anatomical regularization and improved stability. More analyses on Gaussian primitives, covariance parameterization, TOF modeling, and TV regularization will be future work.
5.Clarity, visualization, reproducibility, and ethics (R1, R2, R3): The MRI/CT-guided kernel prior does not transfer anatomical intensity into PET; it regularizes the Gaussian tracer distribution within a plausible anatomical manifold. Clinical PET/CT data were retrospectively collected and anonymized, without intervention or identifiable patient information. We also agree that vectorized/high-resolution figures, code release, and more implementation details are valuable and will further address them in future work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 paper proposes a Gaussian Splatting (GS) representation when solving for a low-count PET reconstruction, moving away from discrete grids while favoring interpretability over implicit representations. The method includes a TOF forward model (system matrix, Poisson likelihood) and a kernel-based anatomical prior. The three reviewers acknowledge that the use of GS in the context of PET imaging is novel. Reviewers have highlighted consistent improvements achieved by the proposed method over several baselines (MLEM, DIP, INR) and across brain and thorax datasets, although the ablation study was considered limited. Some issues were raised regarding the lack of discussion of scalability to full-resolution images and in clinical context, computational complexity, dependence on OSEM in the low-count setting, and sensitivity to hyperparameters. These should be discussed in the revised paper. Section 2.2., and the experimental setting (including the number of patients) need clarification. Include vectorized images for the results. Also, provide the code or enough details to guarantee reproducibility.
