Abstract

Structural MRI-to-amyloid PET synthesis has been proposed as a non-invasive alternative for amyloid assessment in Alzheimer’s disease (AD). However, reported performance of identical models varies widely across studies, and increasingly complex architectures have not led to consistent gains. This inconsistency is thought to be caused by a fundamental biological ambiguity: MRI captures neurodegeneration, while PET measures amyloid pathology - two processes that are often temporally decoupled in AD. As a result, similar MRI patterns may correspond to different amyloid states, creating ambiguous one-to-many mappings. MRI-to-amyloid PET synthesis may therefore be intrinsically ill-posed; however, this idea has yet to be tested scientifically. The aim of this work is to test this hypothesis through two controlled experiments. We first control the training distribution by stratifying paired MRI-PET data by amyloid and neurodegeneration status. Using two standard synthesis models under a controlled design, we show that biologically unambiguous mappings are learnable in isolation, but performance collapses when data ambiguity is introduced. This demonstrates that ambiguity in the data distribution, rather than architectural capacity, constrains performance. Second, we show that introducing orthogonal biological information in the form of plasma biomarkers resolves this ambiguity. When multimodal inputs are incorporated, performance improves and stability is restored. Together, these findings suggest that limited and inconsistent performance in MRI-to-amyloid PET synthesis is explained by intrinsic biological ambiguity, and that stable, meaningful progress requires multimodal integration rather than architectural complexity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5073_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/louisebaron/WhenBrainsDisagree

Link to the Dataset(s)

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

BibTex

@InProceedings{BarLou_When_MICCAI2026,
        author = { Baron, Louise E. G. AND Callaghan, Ross AND Cash, David M. AND Weston, Philip S. J. AND Azadbakht, Hojjat AND Zhang, Hui},
        title = { { When Brains Disagree: Biological Ambiguity Underlies the Challenge of Amyloid PET Synthesis from Structural MRI } },
        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

    The paper uses controlled experiments to explore the extent to which an established medical image analysis problem—synthesis of amyloid PET scans from corresponding structural MRI scans—is actually solvable. Starting with a matched MRI-PET data set, they identify ambiguities, i. e. cases where similar MRI scans correspond to quite different PET scans. They show that removing such ambiguities from the training set improves synthesis performance by a lot. They then show that when you take the full data set with all of its ambiguities and then add relevant side information to the training process (i. e. , an alternative measure of amyloid status), effectively removing ambiguities in a different way, you improve synthesis performance. The conclusion is that in some sense this particular problem is fundamentally difficult to solve due to biological ambiguity, and therefore we are unlikely to make much progress on it simply by throwing ever-more-elaborate deep learners at the problem.

  • 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 is exactly the kind of paper MICCAI needs more of. We are forever formulating new problems and publishing a solution to that problem; leading to a cascade of other people trying to publish better solutions later on. No one seems to ask whether the problem was ever solvable to begin with; what the upper bounds of performance are. Answering the solvability question would tell us when new attempts to squeeze better performance out of a new algorithm are unlikely to be fruitful. Because the solvability question is rarely addressed, we waste untold person-hours on algorithms that are unlikely to push the field forward. This paper addresses solvability and that is commendable.

  • 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 only weakness I see is that the presentation of plasma Alzheimer biomarkers as “complementary” information to the amyloid PET scans is confusing. Blood and PET are different, that is true, but both of them are estimates of the same underlying biological construct which is brain amyloid burden. I would prefer the blood markers to be presented as a second measurement of amyloid rather than something that is truly “complementary. ”

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    Sometimes the right work to do is the work that generates discussion. This paper does that, which is great. Keep up this kind of work please.

  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    This is quite different from 90+% of submitted papers, almost all of which propose a specific algorithm to solve a specific problem. This paper is likely to generate significant discussion, and some pushback no doubt, because it is asking whether it is a waste of time to try to solve an established problem at all. We need that kind of discussion.

  • 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 investigates why MRI-to-amyloid PET synthesis has yielded inconsistent and often poor results across the literature. The primary hypothesis is that the fundamental problem is biological ambiguity: MRI captures neurodegeneration while amyloid PET measures amyloid pathology, and these two processes are temporally decoupled in Alzheimer’s disease, creating ill-posed one-to-many mappings.
    • To experiments conducted using ADNI data where
    • In experiment 1 the data is stratified by amyloid (A) and neurodegeneration (N) into concordant (A+/N+, A-/N-) and discordant (A+/N-, A-/N+) subsets. Two distinct architecture models (pix2pix GAN and latent diffusion) trained on these subsets achieved strong in-distribution performance but struggled with others.
    • In experiment 2, the imaging information was conditioned on the plasm biomarkers via FiLM modulation which show performance improvement in both models. This demonstrates that the synthesis limitation can be attributed to missing biological information rather than model architecture.
  • 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 present an important, hypothesis-driven analysis of a popular neuroimaging problem of interest. Rather than proposing yet another synthesis model, it asks a fundamental question about why existing approaches fail and designs controlled experiments to answer it. The hypothesis is clearly stated, biologically grounded in established AD progression models, and testable.
    • The stratification into concordant and discordant training subsets cleanly isolates the effect of biological ambiguity. The two distinct model architecture further strengthen the argument for architecture-independence.
    • The large test set (n=1,720) with bootstrap confidence intervals provides statistically reliable estimates.
    • The concordant/discordant models achieving ~82–85% balanced accuracy in-distribution but collapsing out-of-distribution is strong evidence for the ambiguity hypothesis.
    • The paper is well-written with a logical flow from motivation to hypothesis to experiments to conclusions.
  • 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 N-intermediate cases, which likely constitute a substantial fraction of the data, are excluded from the stratified analysis but presumably included in the Baseline. How many subjects fall in the N-intermediate category?
    • For an apples to apples comparison in Experiment 2, it would be helpful to train/test the MRI-only model on exactly the same 589 pairs that have the plasma biomarkers and compare the results.
    • Similarly, the work would also benefit from training a plasma-only prediction model from completeness.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (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 provides a principled explanation for the wide performance variability reported across MRI-to-PET synthesis studies and offers a constructive path forward (multimodal integration rather than architectural complexity). This has practical value for researchers working in this space.

  • 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

    The content of this work focus on the usage of image generator, including GANs and diffusion models, to derive PET-amyloid images. The authors demonstrate that brain MR images are limited in generating synthetic PET-amyloid images, since some patients do not display visible anatomical neurodegeneration (detectable by MRI) associated with the deposition of amyloid plaques. To overcome this, they showed that by combing plasma markers related to amyloid plaques, it is possible to derive more accurate and reliable synthetic PET-amyloid. The experimental setup planned by the authors proves the point that spatial information derived from MRI may not be sufficient to derive accumulation of amyloid where neurodegeneration is not visible, regardless of the architecture and methods used.

  • 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 presented a well-organized experimental setup to prove their research questions. The concept of combining biomarkers and MRI images is not novel but it serves well the purpose of directing image generator to derive accurate images.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Some concepts presented are a little redundant. It is a bit questionable whether it is then needed to synthesize PET images when, maybe your information can be maybe derived by the blood markers already.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (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 setup of expriments presented by authors and prove their research question. The use of image generator while combing biomarkers is an interesting approach to obtain surrogate of anatomical and molecular information, but require further investigations.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

We sincerely thank the reviewers for their very positive and constructive feedback. We are particularly pleased that the reviewers found the hypothesis and experimental design regarding the solvability of MRI-to-amyloid PET synthesis meaningful and well-supported. We also appreciate the reviewers’ valuable suggestions. We summarise our responses below.

R1: “the presentation of plasma Alzheimer biomarkers as “complementary” information to the amyloid PET scans is confusing” Response: We agree that the wording could be clarified. Our intention was not to describe plasma biomarkers as complementary to amyloid PET, since both reflect the same underlying biological process. Rather, we meant that plasma biomarkers provide information complementary to structural MRI. We will revise the wording accordingly.

R2.1: “How many subjects fall in the N-intermediate category?” Response: A total of 447 subjects (36%) in the Baseline training and 672 subjects (39%) in the full test set fell into the N-intermediate category. These cases lie between the two neurodegeneration extremes and therefore contribute to ambiguity in a more continuous manner than our defined N-/N+. We verified that including or excluding these cases does not change the overall behaviour or conclusions of the study. Nevertheless, we retained them because they represent a substantial and clinically realistic portion of the data.

R2.2: “in Experiment 2, it would be helpful to train/test the MRI-only model on exactly the same 589 pairs that have the plasma biomarkers and compare the results” Response: We agree this would provide a more direct comparison and will consider it in future work, though it would substantially reduce sample size and may affect statistical power and stability.

R2.3: “the work would also benefit from training a plasma-only prediction model from completeness” Response: We appreciate this suggestion and agree that such an experiment would be informative. Our primary goal with Experiment II was not to demonstrate the superiority of multimodal prediction in itself, but rather to test whether it could reduce ambiguity. However, we agree that a plasma-only model would generally help to directly quantify the relative contributions of MRI versus molecular biomarker information. We will explore this in future work.

R3: “It is a bit questionable whether it is then needed to synthesize PET images when, maybe your information can be maybe derived by the blood markers already.” Response: This is an important point, and we agree that plasma biomarkers raise a broader question regarding the utility of PET synthesis. In this work, however, the primary role of plasma biomarkers was to test the biological ambiguity hypothesis, rather than to justify the clinical necessity of PET synthesis itself. At the same time, we believe PET synthesis may still provide clinically useful spatial information that is not directly available from plasma alone, and hence remain valuable from an interpretability perspective

Once again, we sincerely thank all reviewers for their valuable feedback and encouraging comments.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    I will follow the decision of all the reviewers. R1 recommended a strong acceptance, while R2 and R3 recommended acceptance. Except of some minor fixable weaknesses the reviewers state that the the work is well-executed and statistically credible.



back to top