Abstract

Simultaneous dual-tracer positron emission tomography (PET) enables the acquisition of complementary physiological information within a single scan. However, separating individual tracer signals from mixed measurements remains a severely ill-posed inverse problem due to complex kinetic coupling and high noise levels in dynamic images. Traditional voxel-wise kinetic fitting suffers from heavy noise amplification and the need to estimate many kinetic parameters, while deep learning approaches rely on paired training data that are rarely available in clinical practice. In this work, we introduce a kinetics-informed neural representation framework for dual-tracer separation. Instead of estimating kinetic parameters independently for each voxel, we use a neural network that predicts tracer kinetic parameters from spatial coordinates. This shared representation introduces implicit spatial regularization, enabling spatially coupled and more robust parameter estimation. To realize this, the spatial coordinates are encoded using hash encoding and mapped to physiologically constrained kinetic parameters through an MLP. Then the predicted kinetic parameters are used in compartmental kinetic models to reconstruct separated dynamic images for each tracer. Experiments on both synthetic and real patient data demonstrate that the proposed method achieves superior separation performance compared with conventional and deep learning approaches, highlighting the potential of INR-based kinetic modeling for dual-tracer PET separation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1699_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{WanXin_KineticsInformed_MICCAI2026,
        author = { Wang, Xingyue AND Ye, Qing AND Liang, Ganglin AND Lu, Yihuan AND Cui, Zhiming},
        title = { { Kinetics-Informed Neural Representation for Dual-Tracer Dynamic PET Separation } },
        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 introduces a hybrid framework that integrates implicit neural representations, compartmental kinetic modeling, and case-specific optimization for dual-tracer PET separation. The framework utilizes hash encoding with a MLP to convert 3D spatial coordinates into tracer kinetic parameters. These parameters are then used to generate TACs and dynamic PET images. The optimization process has three terms, including data fidelity to the measured dual-tracer TACs, a total-variation regularizer, and a blood-volume constraint. This study includes both synthetic brain data and real patient-derived data. In the synthetic analysis, the proposed method shows visual comparability to the supervised DTS-INN approach and demonstrates lower noise levels than LM, although the authors note some deviations in kinetic parameters, particularly for the more complex second tracer. In the real-data experiment, the authors claim that their method generates cleaner, more accurately separated images and TACs that align better with reference data compared to the LM method, while the DTS-INN approach demonstrates decreased generalization.

  • 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 proposed method is technically interesting to the field. 2.The proposed method avoids the need for paired training data.

  • 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 relies heavily on visual comparisons and a few ROI parameter tables. There is a lac of tother quantitative metric, such as RMSE, SSIM comparison. 2.How do you evaluate the difference of the two injection’s time? Is 5 minutes enough? How about its generalizability? 3.My major question is related to Table I and II. In the FDG results, we could see much lower bias in Vb and K1, while others demonstrates a higher bias. In the MET results, we could see all the parameters other than vb all shows a higher bias. These results may partially reflect that the proposed method can not work well. Here is the reason: based on the basic kinetic modeling, please note that, K1 and Vb estimation’s sensitivity heavily depend on the early frames, as demonstrated in multiple papers. In other words, if you have 5 minutes time difference, you are be able to estimate an accurate K1 and Vb in the FDG case (this is the single-tracer), which is not related to the power of your proposed method. In the MET results, you have different ground truth for vb, which I think it is unreasonable from physiological perspective. Please clarify this clearer.

  • 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.

    (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?

    The method is technically interesting. However, I highly recommend authors to reorganize the whole demonstration part according to physiological knowledge, identifiability study, and rigorous simulation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    In fact, I can get the response about “We agree estimations of K1 & Vb depend strongly on early frames, but k2-k4 are highly coupled, so isolated biases do not necessarily indicate separation failure.” Based on the quantitative results shown in Table 1 and 2, only K1 and vb showed reasonable bias range thanks to the early-time single-tracer TAC. However, all the other K parameters all showed a large percentage bias (>20%), which can not demonstrate the benefits of the proposed method in tracer separation.



Review #2

  • Please describe the contribution of the paper

    The paper proposes an implicit neural representation (INR) framework to estimate voxel-wise tracer kinetic parameters for staggered dual-tracer PET. By replacing independent voxel-wise fitting, which is typically noisy, with a spatially coherent implicit model, the approach provides regularization across the image domain. The method operates directly on dual-tracer dynamic PET data without requiring single-tracer scans for training and jointly recovers kinetic parameters for both tracers, enabling reconstruction of separated single-tracer dynamic PET images.

  • 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.

    Novel INR-based formulation for dual-tracer PET: Uses an implicit neural representation to estimate voxel-wise kinetic parameters, introducing a spatially coherent alternative to noisy independent fitting. No need for single-tracer training data: Operates directly on dual-tracer dynamic images, improving practicality where ground truth single-tracer data is unavailable. Built-in regularization of kinetic estimation: The INR enforces spatial smoothness, addressing instability in voxel-wise kinetic modeling without explicit regularization. Demonstrated on real data: Inclusion of patient data supports clinical relevance beyond simulations.

  • 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 INR formulation is not clearly defined (e.g., patient- vs. tracer-specific). It appears to be patient- and tracer-pair specific, which may require per-scan optimization. However, training time and hardware requirements are not reported, making clinical feasibility difficult to assess.
    • The paper does not describe how noise was modeled or added in the synthetic data.
    • The chosen tracer pair (11C-MET + 18F-FDG) may represent a relatively easier separation scenario due to differences in half-life. Additionally, if the application targets brain tumors, the absence of a tumor in the synthetic data limits realism.
    • Comparisons to standard analytical methods or alternative neural approaches are missing from several tables and figures, making it difficult to assess performance gains.
    • Results lack error bars or statistical analysis, so significance cannot be evaluated.
    • Unexplained hash encoding: Key parameters ((L, T, F, N_{\min}, b)) are introduced without explanation.
  • Please rate the clarity and organization of this paper

    Poor

  • 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
    • Why were different staggered injection times used (5 min synthetic vs. 10 min real data)?
    • Fig. 4 caption has a capitalization typo.
    • Tables 1 and 2 should be moved closer to their first mention. They also lack uncertainty metrics, making it unclear how to interpret performance. Including baselines (e.g., DNN-INR, analytical methods) or using a bar plot could improve clarity.
    • How is the representative case selected?
    • Where is the LM result? It is described as noisy but not shown.
    • Fig. 5: error bars are missing, and the legend is unclear (e.g., whether LM is the reference). Text size should be increased for readability.
    • Fig. 2 and Fig. 3 are redundant; consider consolidating.
  • 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?

    The paper presents a potentially interesting application of implicit neural representations for dual-tracer PET, particularly in addressing noise in voxel-wise kinetic estimation and demonstrating feasibility on real data. However, the overall contribution is limited by several key weaknesses. The INR formulation is not clearly defined, and it appears to be patient- and tracer-pair specific, potentially requiring per-scan optimization, yet training time and hardware requirements are not reported, making clinical feasibility unclear. Important modeling details, including noise generation in synthetic data and the role of hash encoding parameters, are not described, limiting reproducibility. The experimental evaluation is also not sufficiently convincing: the chosen tracer pair may represent a relatively easy case, synthetic data lacks realism (e.g., no tumor), and comparisons to standard analytical or alternative neural methods are missing in several results. Additionally, results are presented without error bars or statistical analysis, making it difficult to assess significance. Taken together, these issues reduce confidence in both the validity and generalizability of the proposed method, leading to a weak reject despite the promising direction.

  • 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 rebuttal addressed several of my main concerns regarding the INR formulation, implementation details, runtime, and noise modeling. The authors clarified that the method is a scan- and tracer-pair-specific optimization framework rather than a pretrained model. They also responded reasonably to concerns about injection timing and physiological interpretation of the kinetic parameters. They also promised to address comments about figures and tables in camera ready version of paper.



Review #3

  • Please describe the contribution of the paper

    The authors describe a deep learning method to separate the signal of two tracers in dual tracer PET imaging. The authors overcome limitations of other works (high noise, need of paired training data) by a neural network predicting kinetic parameters from spatial coordinates encoded via hash encoding leading to robust parameter estimations. With this method the authors propose a clever way to train a neural network for dual tracer 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 authors use hash-encoding to model kinetic parameters effectively capturing details -Compartment models are integrated in the workflow ensuring a plausible relationship between tracer kinetics and tissue activity

    • The combination of hash encoding and a lightweight MLP allows for fast training and inference, making it scalable for high-resolution PET data.
  • 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.

    In my opinion, the major weakness is that neural fields and MLPs are difficult to interpret. While the output (e.g., kinetic parameters) may be physiologically meaningful, understanding why the model predicts specific values for a given voxel can be challenging. Moreover, if the reconstructed images contain artefacts, it is very difficult to assess where the artefacts are coming from. Moreover, the authors compare their work only to one CNN working with 2D-slices (which is likely that it will not perform equally well) but do not compare their approach to other methods which might work equally well. Additionally, the description of the pipeline lacks details, making it impossible to reproduce the proposed approach. Therefore, while the approach is interesting, it lacks comparison to other works (e.g. https://www.mdpi.com/2076-3417/13/7/4089, https://iopscience.iop.org/article/10.1088/1361-6560/acc723).

  • 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

    The description of the whole method lacks clarity: The authors use an MLP but its lacking architecture details making it impossible to reproduce the results. Also a short phrase what the ‘tracer-specific parameters’ (the output of the MLP) are would help to understand the whole pipeline. The loss function is also vague described: L_vb is penalizing values outside the range [0,1]. But how?

  • 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?

    As mentioned above, the method is conceptually interesting. However, there is missing a comparison to already existing similar method making it difficult to assess its true contribution and details in the method description are missing.

  • 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.

    If the authors include a more detailed description of their method, the quality of the paper is satisfactory. However, to be fully convinced metrics of synthesized cases would be nice to know. Additionally, although better the SSIM values of the proposed method are with values around 0.6 rather low.



Author Feedback

We thank R1, R2, R3 and Meta reviewer (M) for all insightful comments. Our responses are organised as below. 1.Method clarification. Our method targets the practical setting where only dual-tracer scans are obtained with paired single tracer images unavailable. Unlike deep learning methods relying on paired single data for training, we directly optimize each case to obtain separated signals from dual-tracer scans under kinetic constraints. 2.Quantitative evaluation (R1,R2,M). On real data, 3D metrics over nonzero frames show that, against DTS-INN, our method improves PSNR by 18.96/14.96 dB and SSIM by 0.684/0.628, and reduces NRMSE from 0.963/0.720 to 0.027/0.069 for FDG/AV45.Against LM, it improves PSNR by 6.31/10.83 dB and SSIM by 0.174/0.420, and reduces NRMSE from 0.061/0.268 to 0.027/0.069.For synthetic data, only dual-tracer measurements were simulated to match real acquisitions, so the GTs in Figs.2-3 are activity templates from noise-free TACs and our evaluation relies on TAC and visual comparison. We will add metric comparisons of TACs and report mean±std across synthetic cases. 3.K1&Vb concern (R1,M). We agree estimations of K1 & Vb depend strongly on early frames, but k2-k4 are highly coupled, so isolated biases do not necessarily indicate separation failure. If the gain came only from early-frame, LM should achieve comparable TACs under the same staggered information. But Fig. 5 shows that LM deviates at later frames while our method stays close to the reference, confirming the effectiveness of ours. For MET Vb, we agree that it should be consistent across tracers and will revise this in later simulations. 4.Injection interval (R1,R2). Injection interval trades off clinical feasibility and separation difficulty. Prior work (Pan, 2023) showed longer delays improve separability. Thus, we used 5 min for synthetic data as a challenging setting and 10 min for real data following the available protocol. Since our framework is case-specific and the input function includes injection timing, it can be extended to other intervals. 5.INR & Implementation (R2,R3,M). Our INR is scan and tracer pair specific, with the network optimized for each dual-tracer scan. As shown in Fig. 1, it does not synthesize images directly but maps voxel coordinates to tracer specific kinetic parameters. INR extends traditional voxel-wise curve fitting to image domain parameter estimation, improving spatial consistency over noise-sensitive fitting algorithms. Thus, interpretability comes from these constrained kinetic parameters rather than MLP. Hash encoding parameter definitions will be clarified in camera-ready. In L_Vb, Softplus ensures non-negativity and ReLU(Vb-1) penalizes Vb>1.Each synthetic case converges in 27.17 min on RTX 3060.6.Further baselines (R3). Direct comparison is limited by data, protocol, output form and code availability. Ref.1 requires paired single tracer data and outputs static images under a specific protocol, making fair comparison difficult. Ref.2 is based on spectral analysis and limited to image separation, so results mainly focus on TAC analysis. Under this setting, our method reduces NRMSE of compared TACs from 0.131/0.226 to 0.039/0.067 for FDG/AV45.7.Tracer pair/noise/tumor (R2). Although FDG and MET differ in half-life, the 5-min stagger still remains challenging. Our method is not limited to this tracer pair and can be extended to other combinations. Synthetic data were generated by Monte Carlo simulation (Lu,2023 & Sun,2023) with noise arising from the acquisition process. While current phantom was designed for controlled validation, we will add tumors in later simulations. 8.Case selection and figures (R2). The synthetic visual case was randomly selected, while the real data visualization used the only available case with same patient, dynamic brain PET of 2 tracers. Caption typos, table placement and Fig. 2/Fig. 3 overlap will be fixed, with DTS-INN TACs added in Fig. 5.Code will be released.




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 introduced a kinetics-informed neural representation framework for dual-tracer separation. Instead of estimating kinetic parameters independently for each voxel, a neural network predicts tracer kinetic parameters from spatial coordinates through hash encoding and physiologically constrained kinetic parameter mapping through an MLP.

    Reviewers find the unsupervised, case-specific optimization technically interesting, and acknowledge its potential to improve spatial coherence in parameter maps. However, significant concerns are raised regarding INR formulation, the limited interpretability of neural fields and MLPs, the physiological validity of the results, and the limited generalization ability on pre-defined tracer pairs. R1 pointed out that the reported kinetic biases may indicate an inability to accurately separate the tracers beyond simple early-frame estimation. Also, the paper could benefit from improving its clarity and reproducibility since there are currently missing critical details including the MLP architecture, hash encoding parameters, and computational cost. The evaluation is currently mostly qualitative and could be insufficient in the aspect of quantitative metrics (e.g., RMSE, SSIM), missing statistical significance testing, and limited comparison against existing state-of-the-art analytical or neural separation methods.

    Thus, the authors are invited for rebuttal to address core concerns from the reviewers and clarify more details in the methodology.

  • 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.

    Thank the authors for replying to all the reviewers. The merit of the work is acknowledged by most of the reviewers and most of the concerns are sufficiently addressed, except for (1) high bias in k2-k4 and (2) somewhat limited generalization ability to other tracer combinations since the INR is tracer specific, depending on the tracer kinetics. Weighing on all of these, I am leaning toward acceptance of the paper to initiate more discussion in the venue, but would not mind if a majority vote of rejection is given here. The authors are encouraged to further clarify the remaining concerns in addition to integrating edits in the rebuttal round to the manuscript.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    While two of the three reviewers opted for acceptance post-rebuttal, I cannot accept this paper as multiple important aspects are not explained and given the fact that no code is provided, I do not see how the method can be reproduced. Also, the novelty is not that high since to reference 24, and the evaluation is not sufficient to prove the method is an improvement beyond the state-of-the-art.

    Major problems:

    • Reproducibility: The authors do not provide code nor promise to do so. This is in general not a problem, but under those circumstances, the requirements to the explanations are high. Sadly, the authors fail to explain multiple crucial aspects of the paper such as: a) how the arterial input function for each tracer is separated in the patient evaluation - this is not trivial, and key in the parameter estimation; not a word is spent on it. b) details on the MLP architecture (reviewer 3), the training process, the hash encoding (reviewer 2), and noise simulation in the synthetic data (reviewer 2) are not provided / sufficiently explained.

    • Evaluation: Reviewer 1 already points out that the biases observed in the parameters hint that the tracers cannot be separated (see post-rebuttal justification of reviewer 1). Also, as reviewers 1, 2 and meta-reviewer 1 mention, the evaluation is mainly qualitative and the curves/images show practically no difference to the two proposed baselines. Numerical comparisons to the baselines are only partially provided in the rebuttal. More importantly, the baseline selection is suboptimal being the method of reference 24 open-source, and other relevant methods not discussed (see reviewer 3).

    • Novelty: The authors claim that there INRs have never been used for dual-tracer kinetic parameter estimation. This is true, but the method proposed in reference 24 is extremely similar and its code is online. As a result, modifying the code to deal with two tracers is almost trivial in times of coding-optimized LLM. The reviewers actually praised the unsupervised character and the personalization with the “potential to improve spatial coherence in the parametric maps” as meta-reviewer 1 highlights, but that was already introduced in reference 24.Moreover, the strengths reviewers 1, 2 and partially 3 highlight are not new in the light of reference 24.Methodologically, the authors use hashing instead of Gaussian Fourier Features, which is a difference, but it is more of a design choice that likely does not impact the result massively. The use of a total variation regularization and a constraint on the range for Vb are standard choices, so also cannot be considered a contribution. In that sense, this paper is a small modification of reference 24 with new data. As a result, in my view, this is not sufficient novelty for MICCAI.



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.

    Although some reviewer concerns remain, I still find this work to have a strong clinical motivation and to be technically interesting



back to top