Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Multi-tracer positron emission tomography (PET), which assesses key neurological biomarkers such as tau pathology, neuroinflammatory, β-amyloid deposition, and glucose metabolism, plays a vital role in diagnosing neurological disorders by providing complementary insights into the brain’s molecular and functional state. Acquiring multi-tracer PET scans remains challenging due to high costs, radiation exposure, and limited tracer availability. Recent studies have attempted to synthesize multi-tracer PET images from structural MRI. However, these approaches typically either rely on direct mappings to individual tracers, or lack distributional constraints, leading to inconsistencies in image quality across tracers. To this end, we propose a normalized diffusion framework (NDF) to generate high-quality multi-tracer PET images from a single MRI through a distribution-guided class-conditioned weighted diffusion model. Specifically, a diffusion model conditioned on MRI and tracer-specific class labels is trained to synthesize PET images of multiple tracers, and a pre-trained normalizing flow model refines these outputs by mapping them into a shared distribution space. This mapping ensures that the subject-specific high-level features across different PET tracers are preserved, resulting in more consistent and accurate synthesis. Experiments on a total of 425 subjects with multi-tracer PET scans demonstrate that our NDF outperforms current state-of-the-art methods, indicating its potential for advancing multi-tracer PET synthesis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4513_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{YuMin_DistributionGuided_MICCAI2025,
        author = { Yu, Minhui and Lalush, David S. and Monroe, Derek C. and Giovanello, Kelly S. and Lin, Weili and Yap, Pew-Thian and Mihalik, Jason P. and Liu, Mingxia},
        title = { { Distribution-Guided Multi-Tracer Brain PET Synthesis from Structural MRI with Class-Conditioned Weighted Diffusion } },
        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 authors present an approach to synthesise multiple PET images of the brain, corresponding to multiple tracers, from a single anatomical T1w MR image. They propose to do so using a diffusion model conditioned on the MR image and on the tracer type. Two variants are described. The first variant reduces the dimensionality of the PET and MR images before the diffusion process while the second directly takes the (downsampled) 3D PET and MR images as inputs.

  • 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 overall approach, i.e. synthesis of multiple PET images by conditioning on the tracer and anatomical image + diffusion + distribution alignment with normalising flow, appears novel.
    • The proposed approach is compared with a substantial number of existing approaches (six).
  • 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.
    • Even though the global approach appears novel, the methodological contributions are quite unclear. The ‘Latent Space Encoding and Decoding’ is quite usual, the time-varying weighting scheme in section ‘Latent PET Generation with CWD’ is not entirely new (even though this specific implementation may be), and I had difficulties understanding what is being done in ‘Distribution Alignment with NF.’
    • The fact that the subjects of both the private and ADNI datasets are not described is a major problem. We do not know whether they are controls or patients, and if patients, what pathology, and we have no demographic information. PET images may vary a lot depending on these parameters so to me it is impossible to make sense of the experiments without this information.
    • The results are not convincing. The proposed approaches, as well as the existing ones, do not generate tau, PBR and PIB images that match the ground truth. Results appear qualitatively better for FDG PET, but even though the images generated by the two proposed approaches appear realistic, they are not subject-specific. The quantitative results do not demonstrate that the proposed approaches outperform existing ones, as claimed by the authors. Except for DDPM, all the approaches compared lead to similar results.
  • 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
    • 18F-T807 is not the most common name for flortaucipir, this could be changed to help the reader.
    • Ref [5] is about the heart, not the brain. Please find a more appropriate reference.
    • Fig 1: The MRI and PET branches seem switched (Multi-Tracer PET points to Z_M and MRI to Z_P^0).
    • Many references are missing about deep learning-based methods to synthesise PET images from structural MRI. Some recent ones: Ou, Z., Pan, Y., Xie, F., Guo, Q., Shen, D., 2025. Image-and-Label Conditioning Latent Diffusion Model: Synthesizing Aβ-PET From MRI for Detecting Amyloid Status. IEEE Journal of Biomedical and Health Informatics 29, 1221–1231. https://doi.org/10.1109/JBHI.2024.3492020 You, S., Yuan, B., Lyu, Z., Chui, C.K., Chen, C.L.P., Lei, B., Wang, S., 2024. Generative AI Enables Synthesizing Cross-Modality Brain Image via Multi-Level-Latent Representation Learning. IEEE Transactions on Computational Imaging 1–13. https://doi.org/10.1109/TCI.2024.3434724 Zheng, X., Worhunsky, P., Liu, Q., Guo, X., Chen, X., Sun, H., Zhang, J., Toyonaga, T., Mecca, A.P., O’Dell, R.S., van Dyck, C.H., Angarita, G.A., Cosgrove, K., D’Souza, D., Matuskey, D., Esterlis, I., Carson, R.E., Radhakrishnan, R., Liu, C., 2025. Generating synthetic brain PET images of synaptic density based on MR T1 images using deep learning. EJNMMI Physics 12, 30. https://doi.org/10.1186/s40658-025-00744-5 Zotova, D., Pinon, N., Trombetta, R., Bouet, R., Jung, J., Lartizien, C., 2025. GAN-based synthetic FDG PET images from T1 brain MRI can serve to improve performance of deep unsupervised anomaly detection models. Computer Methods and Programs in Biomedicine 265, 108727. https://doi.org/10.1016/j.cmpb.2025.108727
  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Even though the paper may present some novelty, the method is not sufficiently well explained to fully understand the contributions and the experiments are not convincing.

  • 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.

    I would like to thank the authors for their rebuttal. I will keep my recommendation of reject based on the following reasons:

    • Limited methodological novelty. Even though the methods bring some novelty, I still find it somewhat limited. The authors for example did not really explain in what their “timestep-dependent loss weighting” is different from works such as Hang et al., ICCV, 2023.
    • Unconvincing results. The results obtained by all the methods (SOTA and proposed) are globally poor. From Fig. 2 the tau, PBR and PIB images generated do not match the ground truth. The quality of the FDG PET is better but the images generated with the method proposed are not subject-specific (Fig. 3). So even though the PSNR, SSIM, MAE and NMI obtained with the proposed methods often reach the best performance compared with the SOTA, the baselines are so low that having better results does not mean having good results.



Review #2

  • Please describe the contribution of the paper

    This paper works on synthesizing multi-trace PET from T1w MRI, whose direct implementation has good clinical value but a practical barrier. The authors proposed a distribution-guided class-conditioned weighted diffusion model.

  • 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 methodology is well thought out and rationalized. Conditioning the diffusion synthesis on PET contrast class is very natural. Both latent and direct imaging-based info flow had been explored, investigated, and compared. Normalizing flow is to provide further but moderate improvement by regularizing the output distribution. All modules are well designed. Results are presented 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.

    Overall high quality. Statistical tests could be performed to indicate the significance of comparisons in Table 2, and facilitate easier interpretation. For example, in the NDF comparison, for PBR and FDG, the results from w.o WL, w.o NF, the one with TC, WL, NF (“ours”) are likely not to present any statistical difference.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not 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?

    It is a well-written paper on par with all previous MICCAI papers. I do hope and expect the authors to disseminate implementation of this work.

  • 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

    This paper introduces a Normalized Diffusion Framework (NDF) for synthesizing multi-tracer PET images from a structural MRI. It combines: a class-conditioned weighted diffusion (CWD) model to generate PET images guided by tracer labels, a normalizing flow (NF) model to align synthesized images with a learned distribution, ensuring subject-specific consistency across multiple tracers. Additionally, an extension (NDF-I) operates directly in image space to better preserve fine-grained anatomical information. Extensive experiments on a private dataset and the public ADNI dataset demonstrate superior performance over multiple state-of-the-art baselines.

  • 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.

    Innovative Methodology: The integration of class-conditioned diffusion models with normalizing flows for distribution regularization is highly original and effective. Dynamic Noise Weighting: Introducing a weighted noise loss tied to diffusion timestep improves training stability and synthesizing accuracy. Strong Experimental Validation: Comprehensive comparisons across six baselines and two datasets with multiple evaluation metrics. Clinical Relevance: Addresses critical challenges of multi-tracer PET imaging such as cost, radiation, and tracer scarcity.

  • 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.

    Lack of Hypothesis Testing: Although the method outperforms baselines, no statistical significance testing (e.g., paired t-tests) is performed, making it difficult to assess the reliability of small performance differences. Limited Cross-Dataset Generalization: Training and testing are dataset-specific, with no cross-center or cross-scanner generalization evaluation (e.g., training on private dataset, testing on ADNI). Inference Speed Not Addressed: The model uses 1000 reverse diffusion steps without exploring faster alternatives (e.g., DDIM, fewer-step sampling), which impacts clinical feasibility. Dependence on Pre-trained NF Quality: The success of the framework heavily depends on the quality of the NF model; no robustness analysis or failure scenarios are discussed. Potential Overfitting Across Tracers: Multi-tracer imbalance (different sample sizes per tracer) could lead to overfitting, especially for underrepresented tracers, which is not analyzed. No Comparison with Newer Diffusion Models: The paper does not compare with recent diffusion innovations like ControlNet or Denoising Diffusion Transformer (DiT), limiting its positioning relative to the state-of-the-art. Evaluation Limited to Image Metrics: Only PSNR, SSIM, MAE, and NMI are reported. No task-specific clinical evaluation (e.g., diagnosis impact, lesion detection accuracy) is provided. Assumption of Perfect Registration: The model assumes that MRI and PET are perfectly aligned; the impact of registration errors is not evaluated. Robustness to Real-World Noisy MRI Inputs Not Studied: Inputs with artifacts, low-field MRI, or noise are not considered, limiting real-world applicability.

  • 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

    How would minor misregistration between MRI and PET affect the results?

    How does your method handle MRI inputs with motion artifacts or lower field strength?

    What strategies could mitigate dependence on NF pretraining quality?

    Can faster diffusion sampling (e.g., 50-100 steps) be adopted without compromising image quality?

    Would including task-specific clinical metrics (e.g., Alzheimer’s staging accuracy) strengthen validation?

    How would the method compare against newer diffusion frameworks like ControlNet or DiT?

    How does the method handle tracer data imbalance during training?

  • 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 presents a highly novel and well-validated framework for multi-tracer PET synthesis, with important clinical implications. Despite some missing analyses (e.g., significance testing, runtime optimization, generalization), the strengths in methodology, innovation, and experimental validation warrant acceptance.

  • 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.

    This paper presents a methodologically sound and timely approach for synthesizing multi-tracer PET images from structural MRI using a class-conditioned weighted diffusion model, combined with a normalizing flow for distribution alignment. The proposed hybrid diffusion-flow framework is novel in its integration of label conditioning, time-dependent loss weighting, and distribution-guided regularization — which is particularly valuable for tracer-specific image synthesis. While there were initial concerns regarding the clarity of methodology and adequacy of evaluation — especially regarding the interpretation of clinical image fidelity across tracers — the authors have effectively clarified these points. They explained the motivation and implementation of the Normalizing Flow module, addressed subject cohort details (noting that only cognitively normal adults were included), and provided justifications for their architectural choices. They also added statistical validation (paired t-tests), explained robustness to NF training, and acknowledged areas for future evaluation such as generalization to unseen datasets and resilience to noisy MRI inputs. The quantitative and qualitative results demonstrate that the method is competitive across multiple tracers, with especially strong performance in the more challenging cases (e.g., Tau and PBR). The approach is modular, well-motivated, and clearly extensible, with relevance to both clinical and research applications. Given the clarification of methodological aspects and the clinical applicability of the work, I find the paper appropriate for acceptance.



Author Feedback

We thank AC and Reviewers for the constructive feedback. We are especially grateful for the recognition of our methodology innovation, clinical relevance, and rigorous evaluation. We address key concerns (especially from R1) below.

  1. Methodological Clarification (R1)
    • Unlike traditional latent diffusion models, our contribution lies in the integration of timestep-dependent loss weighting with tracer-label conditioning. Our method uniquely uses a time-varying weighting loss for denoising and tracer characteristic modeling, improving stability and tracer-specific PET synthesis. Ablation results (w/o TC, w/o WL) in Table 2 show notable degradation, validating our design.
    • Another of our highly novel contributions is distribution alignment with NF. We will include more details on this in final version, including: a) We aim for synthetic PET to look realistic and match true tracer distributions. Prior models do not explicitly enforce distributional alignment. This work introduces a separately trained NF module that learns a bijective mapping from real PET images to a Gaussian latent space, enabling explicit distribution alignment. b) During training, we use NF to evaluate synthetic PET data’s log-likelihoods, penalizing unrealistic outputs (e.g., anatomically incorrect or tracer-inconsistent) with low scores. This guides the model to produce multi-tracer PET images with higher fidelity.
  2. Missing Subject Detail (R1)
    • All subjects are cognitively normal adults. This design isolates physiological variability. Demographics (e.g., age&gender) will be included in the final version.
  3. Results and Quantitative Validation (R1)
    • As reported in Table 1, our methods achieve the best performance on Tau and PBR, two of the most challenging tracers to synthesize. Paired t-test results (omitted due to page limit) confirm improvements over the top 2 competing methods are statistically significant (p<0.05).
    • For FDG, high SNR and global uptake patterns allows strong baseline performance. For PIB, the sparse cortical uptake limits voxel-wise metrics. Future work could incorporate region-aware or radiomic priors to improve synthesis quality.
    • As noted by R4 and R5, our framework is “highly original and effective,” with “good clinical value” and relevance to “critical challenges in multi-tracer PET imaging.” While our method may not lead on all 4 tracers in Table 1, it performs consistently well overall, especially where precise tracer-specific modeling and subject-level anatomical fidelity are essential.
  4. Statistical Test and Generalization (R4,R5)
    • We will perform paired t-test to quantify improvements and explore cross-dataset generalization by training on private data and testing on others (e.g., ADNI, AIBL, and OASIS) in future work.
  5. Robustness Analysis and Practical Considerations (R5)
    • Inference Speed: Despite 1,000 diffusion steps, our NDF runs efficiently at 0.5s per image during inference, which is clinically feasible. We’ll clarify this in the final version.
    • Dependence on NF Quality: Our NDF remain highly robust to NF quality when training exceeds 10 epochs. With NFs trained for 10, 50, 100, 150 and 200 epochs, results on test data are stable over 4 tracers, while variance slightly decreases with increased NF training. It suggests that our NDF is insensitive to NF model quality in terms of image quality, though longer-trained NFs yield more consistent results. We will include this in the final version.
    • Tracer Imbalance: This was addressed by down-weighting FDG loss by a factor of 10 for training. See “Implementation Details”.
    • Misregistration and Noisy MRI Input: Great suggestion! We’ll assess model reliability by testing NDF with misregistration, motion artifacts and low-field MRI inputs in future work.
  6. Comparison with Recent Diffusion Models (R5)
    • We appreciate R5 for highlighting newer methods like ControlNet and DiT, and will compare them in the future to contextualize our approach.

Thank you!



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.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The novelty is limited and the experiments are not convincing. All the subjects are cognitively normal adults; there is no evaluation for brain desease.



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’

    Reviewers note strengths are the novelty of the proposed method and the extensive experimental comparisons. Concerns included lack of statistical testing, evaluation limited to reconstruction metrics and no downstream tasks, and performance of the method. Post rebuttal, two reviewers were very enthusiastic of the work, and one felt that the novelty was limited and results unconvincing due to qualitatively generally low performance by all methods. Still, the proposed method showed improvement over other approaches with some tracers showing real promise, and thus the proposed approach may help move the field in a positive direction. Given the potential interest in the proposed approach, I recommend accept. The authors should please include the requested clarifications/missing details to strengthen the final version.



back to top