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

Abstract

Alzheimer’s disease (AD) is marked by the progressive evolution of amyloid, tau, and neurodegeneration biomarkers. While MRI provides accessible structural information, PET offers molecular specificity but remains invasive and costly, limiting frequent longitudinal monitoring. To address this, we propose Uni-Brain, a unified diffusion framework for multi-tracer PET synthesis and counterfactual prognosis. Here, counterfactual prognosis refers to individualized PET generation under fixed baseline anatomy and user-specified clinical covariates. Uni-Brain leverages a shared MRI-PET latent space and a conditional latent diffusion model to generate target-tracer PET volumes with anatomical fidelity and clinical controllability. The framework introduces (i) anatomy-informed spatial conditioning, which injects MRI latent codes as dense spatial guidance to mitigate structural drift, and (ii) hybrid clinical conditioning, which separately encodes continuous covariates (e.g., age, CDR-SB) and discrete attributes (e.g., APOEε4, sex) for smoother covariate control. Uni-Brain supports both longitudinal prognosis and counterfactual clinical intervention without requiring paired longitudinal training scans. Extensive experiments on the ADNI dataset demonstrate competitive cross-modal synthesis performance and indicate that the generated follow-up PET images faithfully reflect longitudinal disease progression, exhibiting strong agreement with real follow-up PET images and established clinical biomarkers. Code is available at https://github.com/hjYu-711/UniBrain.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1396_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/hjYu-711/UniBrain

Link to the Dataset(s)

https://adni.loni.usc.edu/

BibTex

@InProceedings{YuHon_UniBrain_MICCAI2026,
        author = { Yu, Hongjie AND Yuan, Haoyue AND Sun, Kaicong AND Shen, Dinggang},
        title = { { Uni-Brain: Anatomy-Informed Diffusion for Multi-Tracer Synthesis and Counterfactual Prognosis from Cross-Sectional Data } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a latent diffusion framework to synthesize multi-tracer PET images from sMRI and clinical attributes, aiming to provide a non-invasive alternative to radioactive imaging. By leveraging dense anatomical conditioning and clinical metadata, the model enables counterfactual generation for controllable disease progression visualization. The approach is validated by assessing both image fidelity and the biological consistency of synthesized pathological burdens across various clinical stages.

  • 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.
    • This paper proposes a latent diffusion framework to synthesize multi-tracer PET images from sMRI and clinical attributes, providing a non-invasive alternative to radioactive imaging.
    • The method leverages dense anatomical constraints and clinical metadata to enable counterfactual generation for controllable disease progression visualization.
    • Notably, the inclusion of Gaussian Fourier Feature mapping for continuous clinical covariates effectively addresses spectrum bias, ensuring robust conditioning that is often overlooked in scalar-driven 3D medical image synthesis.
  • 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.
    • Risk of Deterministic Mapping: The use of an $x_0$ objective combined with dense sMRI anatomical conditioning raises a significant concern. Given the strong biological coupling between sMRI and PET, the model might bypass the iterative denoising process and collapse into a direct sMRI-to-PET reconstruction (regression). In that case, the iterative diffusion framework provides no benefit over a standard Image-to-Image translation network.
    • Missing Diversity Evaluation: The manuscript lacks evaluations of generation diversity. A true Diffusion Model should capture the probabilistic distribution of PET signals. Without visualizing sample variance or measuring sample diversity, it is unclear if the model has actually learned a generative mapping or just a deterministic, one-to-one shortcut.
    • [Minor] Limited Paired Training Data and Risk of Mode Collapse: The proposed method requires strictly paired sMRI, multi-tracer PET, and clinical attributes for training. Such comprehensive datasets are scarce and difficult to access in the medical domain, even for ADNI. For data-hungry generative frameworks such as Latent Diffusion Models, this scarcity heightens the risk of mode collapse or overfitting. Without a sufficiently large and diverse dataset, the model may fail to capture the true underlying distribution of functional pathology, instead settling for a narrow, deterministic mapping that lacks the robustness required for clinical generalization.
  • 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 manuscript presents a technically sound latent diffusion framework for multi-tracer PET synthesis. The pipeline’s design is well-motivated, particularly the use of Gaussian Fourier Feature mapping to handle continuous clinical covariates, which effectively addresses the spectrum bias often ignored in scalar-conditioned medical image generation. The inclusion of dense anatomical constraints from sMRI and clinical attributes for counterfactual visualization provides a promising path for controllable disease progression modeling.

    However, the recommendation is tempered by several concerns: Generative Nature vs. Deterministic Mapping: The reliance on an $x_0$ objective under strong anatomical conditioning risks the model collapsing into a deterministic sMRI-to-PET reconstruction, potentially bypassing the stochastic benefits of the diffusion process.

    Evaluation: While the quantitative results (PSNR/SSIM/SUVR SCC) are solid, further evidence of generation diversity and a clearer justification for the diffusion framework over simpler translation baselines would strengthen the work.

    Data Scarcity: The requirement for high-quality, paired sMRI and multi-tracer PET data is a significant hurdle. For a data-eager generative model, this raises concerns regarding mode collapse and the model’s ability to generalize beyond the limited training distribution.

  • 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 author’s rebuttal sufficiently responded to my review comments. Although some potential improvement are promised in future works, which is acceptable given the scope of the current work.



Review #2

  • Please describe the contribution of the paper

    1.Anatomy informed spatial conditioning that injects MRI latent codes as dense spatial guidance into the denoising network to preserve subject specific anatomical structure. 2.hybrid clinical conditioning that separately handles continuous covariates (age) via Gaussian Fourier Projection and discrete covariates (sex, APOEe4) via learnable embeddings, with an availability indicator to handle missing clinical data.

  • 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.Unifying multi-tracer PET synthesis and counterfactual prognosis within a single framework is novel. Existing diffusion-based synthesis methods such as ResDM treat cross-modal mapping as a static T0 to T0 problem, while longitudinal models require paired multi time point data. Addressing both from cross-sectional data in one architecture is a meaningful practical contribution.

  • 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.Counterfactual prognosis is not validated against longitudinal ground truth

    • The paper’s primary novel task, counterfactual prognosis, is evaluated using cross-sectional proxies rather than actual future PET scans. Without comparing synthesized future PET against real longitudinal PET from the same subject, the individualized prognosis claim cannot be substantiated.

    2.Task 2 baselines are insufficient

    • Brain Latent Progression (Puglisi et al., Medical Image Analysis, 2025) and SADM (Yoon et al., IPMI 2023), both cited in this paper, are not compared against despite being directly relevant.

    3.No statistical significance testing

    • No p-values or confidence intervals are reported for any comparison in Table 1.Given the overlap in standard deviations across methods, it is unclear whether reported differences are statistically meaningful.

    4.Ambiguous CDR specification and unsupported citation

    • The manuscript refers to CDR throughout without specifying whether this is the Global CDR or CDR-SB, which is a meaningful distinction in a clinical context. If Global CDR is intended, values of 4 and 6 in Fig. 4 do not exist on the standard scale. If CDR-SB is intended, this should be stated explicitly. Additionally, the citation used to justify treating CDR as a continuous covariate [16] refers to ResDM, which conditions on clinical text rather than discrete CDR embeddings, and does not support the claim being made.
  • 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 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 unified framing for multi-tracer PET synthesis and counterfactual prognosis is novel and the Task 1 synthesis results are competitive against strong baselines. However, several issues place this paper below the acceptance threshold. The primary novel contribution, counterfactual prognosis, is evaluated without longitudinal ground truth despite its availability in ADNI, and outperforming ResDM under the same cross-sectional setup does not substantiate the individualized prognosis claim. The most directly relevant baselines for Task 2, Brain Latent Progression and SADM, are cited in the paper but absent from the comparison. No statistical significance testing is reported, making it difficult to assess whether the performance differences in Table 1 are meaningful given the overlapping standard deviations. Finally, the clinical variable CDR is used throughout without specifying Global CDR or CDR-SB, and the citation supporting its continuous treatment [16] does not back the claim. These issues are identifiable from the current submission and collectively support a weak reject.

  • 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 substantively addresses three of the four weaknesses I raised.

    1) On CDR specification they provide suitable feedback –> This resolves the concern directly

    2) On longitudinal evaluation, the authors clarify that the test set contains 401 baseline subjects with 162 / 322 / 359 follow-up AV1451 / AV45 / FDG PET scans respectively, and that each synthesized future PET is evaluated against the corresponding subject’s real follow-up PET.

    3) On statistical significance, the authors commit to reporting two-sided Wilcoxon signed-rank tests on existing PSNR/SSIM outputs and bootstrap 95% CLs for the cohort-level SCC. –> I take the reported p < 0.05 in good faith pending the camera-ready.

    4) On the Task 2 baseline comparison, the authors argue that BrLP and SADM are same-modality longitudinal models requiring paired multi-timepoint supervision, while Uni-Brain trains on cross-sectional MRI-PET pairs and infers multi-tracer PET from a single baseline MRI. The training paradigms genuinely differ, but it does not fully eliminate the comparison gap, since both works are cited in the manuscript and the conceptual relevance to longitudinal PET prediction is clear. I would have preferred an explicit justification in the paper rather than omission. However, given that the authors have substantively addressed the other three weaknesses and the conceptual contribution is genuinely novel, I am willing to accept this as a limitation rather than a fatal flaw.



Review #3

  • Please describe the contribution of the paper

    This paper proposes Uni-Brain, a unified conditional latent diffusion framework for multi-tracer PET synthesis and counterfactual prognosis from a single baseline MRI. The method is motivated by the limitation that PET imaging is costly and unsuitable for longitudinal acquisition, while existing MRI-to-PET translation approaches are restricted to static T0→T0 mappings and cannot model disease progression. To overcome this, the authors introduce a shared MRI–PET latent space in which different PET tracers are modeled as sub-manifolds, and where subject-specific anatomy and clinical states jointly determine the latent representation. The framework conditions PET generation on three factors: MRI-derived anatomical latent codes, clinical covariates such as age and CDR, and tracer type, enabling unified modeling across modalities. The model employs a hybrid conditioning strategy that separately embeds continuous and discrete clinical variables to allow smooth and controllable extrapolation. Based on this formulation, the model supports both multi-tracer synthesis via modality switching and counterfactual prognosis via intervention on clinical variables.

  • 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 paper addresses an important and practical problem in neuroimaging, namely the difficulty of acquiring longitudinal PET scans due to cost and invasiveness. By aiming to infer multi-tracer PET and disease progression from a single MRI, the work is well-motivated and aligned with real clinical constraints. 2.The proposed method is validated on a relatively large-scale multimodal dataset with multiple PET tracers, including thousands of MRI and PET scans, which strengthens the credibility of the empirical results.

  • 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 framework lacks a clear formulation of longitudinal modeling, making the claim of “counterfactual prognosis” questionable. The model does not learn explicit temporal dynamics or predict future clinical states, but instead relies on manually specified variables such as age and CDR as conditioning inputs. As a result, it effectively performs conditional image generation under assumed future conditions rather than true forecasting. The manuscript does not clarify how future clinical covariates are obtained or whether they are predicted or externally provided. This ambiguity weakens the validity of the longitudinal setting and raises concerns about whether the model truly captures disease progression over time. 2.The paper claims that hybrid clinical conditioning enables disentangled and interpretable control, but this is not well supported by the model design. The approach simply embeds and concatenates variables and injects them via AdaGN, which does not guarantee disentanglement. There is no explicit constraint or objective enforcing independent factorization of clinical variables. As a result, the method is closer to standard conditional generation, making the disentanglement claim overstated. 3.The unified manifold perspective is conceptually appealing, but not rigorously supported by the model design. In particular, the assumption that different tracers lie on distinct sub-manifolds and that anatomy and clinical variables uniquely determine subject coordinates is not explicitly enforced or validated.

  • 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 paper tackles a clinically important problem, with promising results on a relatively large dataset. However, while the idea is interesting and the problem is important, the gap between the conceptual claims and the actual technical formulation and validation limits the strength of the contribution.

  • 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 appreciate the reviewers for acknowledging our contributions and providing valuable feedback. We provide the responses as follows. R1: 1)Deterministic mapping. x0-prediction is a standard diffusion parameterization recommended for few-step samplers (Salimans & Ho, ICLR 2022) and is mathematically equivalent to ε-prediction up to loss re-weighting under the VP-SDE; inference uses 50-step DDIM from Gaussian noise, not single-pass mapping. The tracer-specific responses under fixed anatomy in Fig.2 provide qualitative evidence against deterministic mapping. 2)Diversity. Our goal is not unconstrained PET generation, but conditional PET synthesis given subject MRI, tracer, and covariates. In this setting, desirable variability should be biologically constrained; excessive diversity may indicate poor conditioning rather than better modeling. We will analyze diversity/uncertainty in future studies. 3)Data scarcity. We agree that limited data increases mode-collapse risk. Uni-Brain mitigates this by avoiding complete MRI-FDG-Aβ-Tau tuples: any same-visit MRI-PET pair contributes via the tracer indicator, and missing clinical covariates are handled by the availability indicator. ADNI-only validation cannot fully rule out overfitting, but Table 1 and Figs. 2/4 do not indicate a trivial averaged PET mapping. We will temper generalization claims and discuss external validation. R2: 1)CDR. The variable used in experiments is CDR-SB, not Global CDR, we will revise CDR to CDR-SB. Our view is that AD severity is better modeled as an continuous variable than as a discrete label, we will replace [16] with Schmidt-Richberg et al., PlosOne and Jack et al., Alzheimer’s & Dementia. 2)Longitudinal evaluation. We apologize for the under-described protocol. The test set contains 401 baseline subjects and 162/322/359 follow-up AV1451/AV45/FDG PET. So each synthetic future PET is evaluated against the subject’s real follow-up PET. We will state this protocol explicitly in the revision. 3)Task 2 baselines. BrLP/SADM are same-modality longitudinal models requiring paired multi-timepoint supervision. Our model trains on cross-sectional MRI-PET pairs and infers multi-tracer PET from one baseline MRI; direct adaptation of BrLP/SADM would require redesigning model architecture. 4)Statistical significance. From existing PSNR/SSIM outputs, two-sided Wilcoxon signed-rank tests give p<0.05 for the all comparisons. Since SCC is cohort-level, we will add bootstrap 95% CIs in the revision. R3: 1)Counterfactual prognosis formulation. Uni-Brain is not autonomous forecasting with explicit temporal dynamics. It follows cross-sectional disease progression modeling paradigm (Young et al., Nat. Commun.; Jedynak et al., NeuroImage), where population-level progression regularities are learned from cross-sectional cohorts and applied to individual baseline feature. We use a unified function F: (s, c, m) → x_pet in two settings: (1) longitudinal progression (Task 2): fix anatomy/tracer, set age to the real follow-up age, and mask future CDR-SB; (2) counterfactual simulation (Fig.4): fix anatomy/tracer and manually specify clinical covariates. 2)Hybrid conditioning. We agree that no independence loss enforces explicit disentanglement. Our goal is covariate-specific processing: Fourier features encode continuous/ordered variables such as age and CDR-SB for smooth control, while class embeddings represent categories such as sex, APOEε4 for implicit classification. 3)Manifold view. We agree that the manifold assumption are not explicitly enforced or rigorously validated. We use the manifold view as a hypothesis-guided design principle, motivating the shared VAE, tracer indicator, anatomy anchor, and clinical conditioning. Our results confirm the empirical validity of this design but do not prove this assumption, we will soften the related claims.



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 proposes Uni-Brain, an anatomy-informed diffusion model for multi-tracer synthesis and counterfactual prognosis from cross-sectional data. R1 finds the method technically sound but has concerns about deterministic mapping, missing diversity evaluation, and data scarcity. R2 acknowledges the novel framing but criticizes the lack of longitudinal validation, insufficient baselines, no statistical testing, and ambiguous CDR specification. R3 appreciates the clinical importance but finds the conceptual claims exceed the technical formulation and validation.

  • 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 authors provide convincing clarifications on the longitudinal evaluation protocol (including the use of subject-specific follow-up PET as ground truth), on the precise definition and use of CDR-SB, and on the statistical significance of their results, directly addressing the major methodological and clinical concerns raised in the initial reviews. Remaining issues, such as the lack of explicit temporal dynamics, limited analysis of generative diversity, and the absence of some longitudinal baselines, are clearly acknowledged, with claims appropriately softened and framed as limitations and directions for future work rather than overstated contributions. Given the clinical motivation, technically sound framework, competitive performance on a challenging multi-tracer setting, and the fact that two reviewers now recommend acceptance while the remaining concerns are not critical, the paper can now be accepted.



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.

    I recommend acceptance. This paper presents a novel and well-motivated framework for multi-tracer PET image generation using a unified conditional diffusion model. The method incorporates structural MRI latent representation, clinical variables, and tracer-specific conditioning, allowing the model to synthesize different PET tracers and generate prognostic and counterfactual images by manipulating the conditioning variables at inference time.

    One of the main strengths of the work is its ability to perform longitudinal prognosis despite being trained only on cross-sectional data. By modifying the age condition and masking the CDR condition, the authors show that the model can generate plausible prognostic PET images without relying on paired longitudinal scans for training. The proposed approach achieves superior performance compared with relevant baselines in both cross-modal synthesis and longitudinal prognosis experiments.

    The authors also provided a satisfactory rebuttal, addressing most reviewer concerns, including the addition of statistical significance testing and clarification of the evaluation protocol. Overall, the paper is technically solid and offers a useful contribution to PET image synthesis and prognosis.



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 rebuttal successfully addressed concerns from R2 regarding several evaluation issues. With the initial positive reviews from other reviewers, I recommend accepting the paper.



back to top