Abstract

Amyloid Positron Emission Tomography (PET) plays a critical role in the early diagnosis and staging of neurodegenerative disorders; however, its widespread clinical use is constrained by high costs, radiation exposure, and limited accessibility. Synthesizing amyloid PET from routine structural Magnetic Resonance Imaging (MRI) represents a promising alternative, yet existing methods often yield unreliable, physiologically inconsistent results due to the lack of prior knowledge about amyloid plaque distribution. To address the problem, this work explores population-level amyloid pathology priors with explicit modeling of spatially varying voxel-wise variability and proposes a fully 3D Template-Guided Heteroscedastic Diffusion Bridge (TGH-DB) for brain MRI-to-PET synthesis. The proposed framework introduces a population-level amyloid pathology template, defined by voxel-wise mean and standard deviation estimated from the training cohort, to steer the diffusion process toward an empirically plausible target distribution. Unlike conventional diffusion models that drift toward an isotropic Gaussian prior, our forward process progressively guides PET signals toward the population-level amyloid pathology template with injection of spatially adaptive, heteroscedastic noise, thereby explicitly modeling regional variability in amyloid uptake. To effectively leverage structural information from various MRI modalities, we further develop the Time-Aware Evolutionary Denoiser (TAED) that integrates multimodal MRI information. TAED comprises parallel encoders, Time-Aware Fusion (TAF) modules, and a time-modulated decoding pathway so that different types of MRI information are adaptively exploited during the diffusion process. Extensive experiments on two public datasets demonstrate that the proposed method consistently outperforms representative existing approaches in both quantitative metrics and qualitative visual evaluations.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0673_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/PangHaowen-hub/TGH-DB

Link to the Dataset(s)

N/A

BibTex

@InProceedings{PanHao_TGHDB_MICCAI2026,
        author = { Pang, Haowen AND Zhu, Yitao AND Fu, Yingji AND Ye, Chuyang AND Qiu, Anqi},
        title = { { TGH-DB: Template-Guided Heteroscedastic Diffusion Bridge for Brain MRI-to-PET Synthesis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a template-guided diffusion framework for MRI-to-PET synthesis, where the conventional isotropic Gaussian endpoint is replaced by a population-level amyloid pathology template characterized by voxel-wise mean and standard deviation maps. It further introduces heteroscedastic noise to encode spatially varying population variability and a TAED-based time-aware multimodal MRI fusion mechanism for denoising.

  • 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 work is conceptually coherent and addresses a real weakness of standard diffusion formulations for medical image synthesis, namely the lack of anatomy- and pathology-aware priors. The proposed template-guided endpoint is the most interesting aspect of the paper and gives the method a clearer medical motivation than simply scaling up the denoiser. The methodology is well structured, the paper is generally easy to follow, and the experimental results are reasonably consistent across datasets and tracers. The ablation studies also suggest that the main components contribute meaningfully.

  • 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’s strongest claims are somewhat ahead of the evidence: the experiments show improved synthesis quality, but do not directly validate the claimed physiological plausibility of the generated PET images. 2.The proposed heteroscedastic design appears to capture cohort-level spatial variability rather than truly calibrated subject-level uncertainty, so the uncertainty-related claim may be overstated. 3.A key assumption of the method—that voxel-wise SUVR distributions approximately follow Gaussian patterns—is asserted but not sufficiently supported by direct statistical evidence. 4.Because the pathology template is derived from the training cohort, the method may introduce cohort-prior bias, potentially pulling rare but real individual abnormalities toward the population mean. 5.Although the paper compares against several methods, it does not clearly include some of the most recent and most closely related MRI-to-PET or distribution-guided baselines, which weakens the strength of the SOTA claim.

  • 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.

  • 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?

    I lean Weak Accept because the paper is logically sound, the main idea is meaningful, and the code release is a positive factor. However, the manuscript currently over-claims some findings, and the SOTA claim is weakened by missing strong recent baselines. So this is only marginally above the acceptance threshold and depends on the rebuttal. These are important but fixable issues, and if the authors address them well in the rebuttal or revision, I would be quite supportive of acceptance.

  • 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 3D diffusion bridge framework for synthesizing brain amyloid PET images from multimodal MRI. Unlike conventional diffusion models that target Gaussian noise, TGH-DB guides the diffusion process toward a population-level amyloid pathology template derived from training data statistics.

  • 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.Using template-based priors instead of standard Gaussian priors is an effective and novel method to reduce the search space in medical image synthesis.

    2.The author proposed voxel-wise variance to capture local variations in amyloid uptake.

  • 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 proposed method relies on a population-level template (μ_tpl, σ_tpl) estimated from the training cohort. However, the paper lacks an analysis of how this model performs on completely independent external datasets where the PET intensity distributions might differ. Specifically, I would like to see the distribution gap between the training and test splits within the current datasets.

    2.The Time-Aware Fusion (TAF) module and the multi-encoder design add significant complexity compared to simple channel-wise concatenation. While the ablation study shows a drop in performance without TAED, the architectural rationale specifically why individual Swin Transformer encoders are necessary for each modality is not sufficiently detailed.

    3.While amyloid-positive/negative classification was performed for the A4 dataset, it was omitted for the HABS-HD dataset, which reportedly includes MCI and AD subjects. Furthermore, classification performance is not reported in the Ablation Study.

    4.The paper states that the template is estimated from the training cohort but does not clarify the composition of this cohort. Since amyloid distribution differs significantly between cognitively normal individuals and those with Alzheimer’s disease, using a single unified template might be suboptimal. It is unclear if the authors used only healthy controls for the template or a mixture of subjects.

  • 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.

  • 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 idea of Template-Guided diffusion is a significant contribution to the field. However, there are concerns regarding the statistical robustness of the template and the motivation for architectural design choices. If the authors can address the concerns regarding external validation and the impact of disease-specific distributions on the template, I would be willing to increase my score.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    The authors have partially addressed my concerns. While the paper is on the borderline, the idea and contributions are strong. I recommend Accept. I hope the remaining concerns raised by the reviewers are addressed in the camera-ready version.



Review #3

  • Please describe the contribution of the paper

    This work explores brain MRI-to-PET synthesis and introduces a population-level amyloid pathology prior, characterized by the mean and standard deviation estimated from the training cohort, to modify the zero-mean, homoscedastic variance setting in the original diffusion process. The method is validated on two public datasets and demonstrates clear improvements in quantitative metrics compared with competing methods.

  • 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.Redefining the mean and variance of the forward noising process is a novel and interesting design, which differentiates this work from conventional diffusion-based synthesis methods. 2.The experimental comparison is relatively comprehensive, and the proposed method shows clear improvements in quantitative metrics over competing approaches. The inclusion of downstream-task validation further strengthens the evaluation. 3.The ablation study is comprehensive and clearly demonstrates the contribution of each component to the overall performance improvement. 4.The paper is well organized and clearly written, with a logical presentation of the motivation, method, and experiments.

  • 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 term “Diffusion bridge” seems not fully appropriate here. The proposed method is still essentially a conditional diffusion model with MRI as the conditioning input. Its main modification lies in redefining the mean and variance of the forward noising process, rather than constructing a bridge process that interpolates between two endpoint distributions, as in diffusion bridge formulations. 2.Public datasets used in this study may exhibit bias in the distribution of amyloid PET phenotypes. For example, the A4 cohort mainly includes cognitively unimpaired individuals with amyloid-positive PET, whereas LEARN and HABS-HD include amyloid-negative subjects and/or participants with different cognitive statuses. The authors should further clarify the proportions of amyloid-positive and amyloid-negative subjects in each training and testing subset, and assess whether such distributional bias may cause the model to favor synthesized PET images with elevated cortical uptake or SUVR. In particular, for MRI inputs from amyloid-negative or non-AD individuals, additional quantitative evaluations are recommended, such as SUVR bias in the negative group, specificity, and false-positive rate, to verify that the model does not systematically generate falsely “positive-like” PET patterns. 3.Normalizing PET images to a standard space facilitates the construction of a population-level template and enables cross-subject comparability of µ_tpl and σ_tpl. However, the required affine and nonlinear transformations may alter individual anatomy. As such, standard-space results are more appropriate for population-level modeling than for direct individual-case presentation. The authors should clarify whether the final output can be mapped back to the original space to support interpretability and practical usability at the subject level.

  • 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.

  • 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 paper addresses a meaningful problem and presents a reasonably novel method. The overall results are promising, and the main idea is well supported. However, a few aspects still require further clarification or additional experiments.

  • 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.

    After reading the authors’ rebuttal, my final recommendation is Accept. The rebuttal has sufficiently addressed my main concerns.



Author Feedback

  • Over-claimed physiological plausibility and uncertainty modeling (AC, R1) We will clarify “physiological plausibility” as empirical spatial distributions and “uncertainty” as voxel-wise variability.

  • Template bias and generalization (AC, R1, R2, R3) Our template is used as priors to stabilize ill-posed estimation problems. It is not the sole factor determining the estimation result, where observed data also plays a substantial role in the estimation. Moreover, we used a large number of PET images for template construction, which better represents the population with reduced bias. Finally, we confirm that templates constructed from different cohorts are consistent. The two templates estimated from A4 and HABS-HD show highly similar patterns, with elevated values in posterior cingulate/precuneus, frontal, lateral temporal, and parietal cortices, and lower values in the cerebellar cortex. This suggests that the template captures a stable amyloid-related spatial pattern rather than dataset-specific artifacts.

  • Template cohort composition (AC, R2) The template is estimated from all training subjects. In future work, we will construct separate normal and abnormal templates and enable their adaptive integration.

  • Classification validation and ablation (AC, R2) We performed amyloid-positive/negative classification on A4 because it is directly aligned with the goal of amyloid PET synthesis. In contrast, MCI/AD is a broader clinical diagnostic task influenced by multiple factors. The classification task was not included in the ablation study because its performance is affected by the classification method, making it harder to attribute performance changes specifically to the ablated component. We will clarify this.

  • Amyloid status bias (AC, R3) In A4, the positive proportions in training/validation/test sets are 69.52%/73.03%/68.63%, respectively. TGH-DB achieves PSNR of 31.51/31.63 and SSIM of 92.88/93.19 for amyloid-positive/negative subjects, respectively, indicating stable performance for both subgroups. Moreover, the downstream classification results support that the model does not systematically generate false positive PET patterns.

  • Gaussian SUVR assumption (AC, R1) In our preliminary analysis, we randomly sampled a large number of voxels and confirmed that their SUVR distributions show Gaussian patterns. We agree that this should be interpreted as an approximation rather than a strict distributional claim.

  • Baselines and SOTA claim (R1) Some recent methods were not included because their task settings differ from ours or code is unavailable. For example, [13] relies on cerebrospinal fluid biomarkers, while [29] synthesizes multi-tracer PET images. We will soften the SOTA claim and state that our method outperforms representative existing methods.

  • TAF and multi-encoder design (AC, R2) The multi-encoder and TAF designs are motivated by the complementary roles of T1W and FLAIR. T1W provides structural contrast and FLAIR is sensitive to anomalies. Independent encoders allow the model to preserve modality-specific representations before fusion. Besides, different diffusion time steps require different MRI cues. Therefore, TAF uses time embedding to adaptively modulate the contribution of each modality. The ablation results indicate the benefits of the design.

  • Diffusion Bridge (R3) We agree that our method is not a classical diffusion bridge in the strict sense. We use the term “bridge” because the forward process introduces a mean drift term and a heteroscedastic noise term to connect PET images with the template. This directed transition and reverse interpolation among the current latent state, the predicted clean PET, and the template make our formulation resemble a bridge-like diffusion process. We will clarify that our method is bridge-like.

  • Native-space usability (AC, R3) The spatial transformation is saved for each subject, and synthesized PET can be mapped back to the native space accordingly.




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 paper received 2 weak accept and 1 weak reject. Reviewers agree that this is a novel and well-motivated idea, particularly in introducing pathology-aware priors into diffusion modelling, and the method shows consistent quantitative improvements with solid ablation studies.

    Key issues to address in rebuttal include: (1) over-claimed conclusions about physiological plausibility and uncertainty modeling (R1); (2) template bias and generalization concerns, including cohort composition and robustness to distribution shifts (R1, R2, R3); (3) discussion on the limited validation (R2, R3); and (4) method clarity and justification, statistical assumptions (e.g., Gaussian SUVR)(R1, R2, R3).

    The authors are encouraged to provide a detailed response to reviewers’ concerns.

  • 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.

    This is a strong retuttal. Reviewers’ main concerns are fully addressed.



Meta-review #2

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

    Accept

  • Please justify your recommendation.

    The rebuttal addressed the main concerns about template bias, architectural motivation, and terminology, and both initially borderline reviewers updated their recommendations to accept. The paper now has a clear consensus toward acceptance, with a novel and well-motivated template-guided diffusion formulation and solid empirical results, while some limitations in physiological interpretation and external validation remain to be refined in future work.



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.

    The submission proposes a novel and well-motivated diffusion framework for MRI-to-PET synthesis. Following the rebuttal phase, the initial skepticism has been successfully resolved, leading to a clear consensus for acceptance from all reviewers. In their response, the authors effectively mitigated concerns from the reviewers. The authors must ensure that all clarifications provided in the rebuttal are fully integrated into the final camera-ready version.



back to top