Abstract

Alzheimer’s disease (AD) is a progressive and irreversible brain disorder. Emerging evidence suggests that Aβ deposition in the heart and microbiota dysbiosis in the gut may also contribute to the pathogenesis of AD. However, currently no studies have integrated heart and gut imaging information into AD diagnosis. To address this gap, we propose the first framework to integrate brain, heart, and gut information based on whole-body PET imaging and leverage these multi-organ interactions to guide brain-only model for early AD diagnosis in clinical applications. To this end, we collect multi-cohort data, including 1,475 unlabeled whole-body FDG-PET images, 1,730 brain FDG-PET images, and 70 labeled high-quality whole-body FDG-PET images. Our AD diagnostic model consists of two stages: (1) feature extraction and alignment, where AD-related features across brain, heart, and gut are extracted and aligned via hierarchical Transformers using contrastive learning; and (2) multi-constraint knowledge distillation, which utilizes sample-level contrastive distillation, group-level distribution distillation, and responselevel distillation to transfer the performance of brain-heart-gut model to the brain-only model. Experimental results show that, guided by the learned interactions of brain, heart, and gut, our brain-only model improves the area under the receiver operating characteristic curve (AUC) from 75.4% to 80.3% for normal control vs. mild cognitive impairment (MCI) classification, achieving comparable diagnostic performance of using whole-body PET.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2631_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lifan0321/BHG-distillation

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiFan_BrainHeartGut_MICCAI2025,
        author = { Li, Fan and Zhao, Shilun and Bai, Shuwei and Liu, Yuxiao and Zhang, Kai and Xu, Yin and Zhang, Ya and Sun, Kaicong and Shen, Dinggang},
        title = { { Brain-Heart-Gut Guided Multi-Constraint Knowledge Distillation for Early Alzheimer’s Disease Diagnosis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15974},
        month = {September},
        page = {44 -- 54}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    In this work, the authors aim to leverage whole-body FDG PET images, with a specific focus on the heart and gut, jointly with brain FDG PET, for the computer-aided diagnosis of mild cognitive impairment. They propose a three-stage approach. The first stage consists in pre-training three independent encoders, each dedicated to a body part. The heart and gut encoders are pre-trained using a reconstruction task and the brain one using a diagnostic classification task. The second stage consists in aligning the features obtained from the three encoders to form a joint brain-heart-gut model. This is done thanks to contrastive learning and both self- and cross-transformers. The third and final stage consists in distilling the brain-heart-gut model to a brain-only model. This is achieved thanks to a three-level distillation framework that includes sample-level contrastive distillation (brain-heart-gut and brain models from the same subject should be similar), group-level distribution distillation (brain-heart-gut and brain models from subjects with the same diagnosis should be similar), and response-level knowledge distillation (brain-heart-gut and brain models should overall be similar).

  • 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.
    • Exploiting whole-body FDG PET images for the computer-aided diagnosis of dementia seems novel.
    • The whole approach appears original and interesting.
    • The experiments assess the influence of the different components of the framework on the classification performance.
  • 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 authors do not really discuss how feasible such approach would be in practice, i.e. whether it is realistic to acquire whole-body PET images routinely.
    • There is no comparison with approaches that handle missing data at test time, and with approaches that could be trained on brain+heart+gut data but applied on brain-only data.
    • The generalisability study is quite weak as data from ADNI and Hospital C were used at stage I, so these are no real external sets.
    • It is not clear from the results if the gain in classification performance with the proposed approach compared with only using brain data is clinically relevant.
  • 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.

  • 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
    • There seems to be mistakes in eq. 2 as not all losses appear in the formula and some appear twice.
    • Please be clear whether a simultaneous PET/MR scanner was used to acquire the data or whether PET and MR images were acquired separately. The term is currently confusing.
    • In the tables, please specify what the standard deviations correspond to (i.e., were they computed across the folds?). Please comment on the fact that the std is sometimes quite large.
    • Please do not use the word “significant” without a statistical test to back it up.
  • 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 — could be accepted, dependent on rebuttal

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

    The method is quite interesting and appears novel, the paper is quite clear, but the clinical relevance is not evident.

  • 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 makes two key contributions (1) It demonstrates that features extracted using self-supervised learning from unlabeled whole-body FDG PET data, specifically from the heart and gut can be meaningfully linked to features from brain data. (2) It is notable for its effort to identify imaging biomarkers for Alzheimer’s disease by integrating information from brain, heart, and gut regions.

  • 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.

    One of the major strengths of this work lies in its technically interesting approach to leveraging unlabeled data. The study applies self-supervised learning to extract features from heart and gut regions using whole-body PET-CT images, and then employs contrastive learning to establish meaningful associations between these features and those derived from brain imaging. This represents a novel and creative use of deep learning methods in the context of multi-organ analysis.

  • 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.

    One major weakness of the paper is the lack of clarity regarding whether the features extracted from brain, heart, and gut images are sampled from the same individual cases within the training set. If these organ-specific features were randomly mixed across different subjects without maintaining case-level correspondence, the contrastive learning in Stage 2 may be unintentionally biased, artificially enforcing correlations that do not exist in reality. Therefore, it is important to first validate whether such bias exists in the intermediate representation before drawing conclusions through the knowledge distillation process in Stage 3.

  • 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.

  • 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.

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    This paper is dependent on rebuttal because it is necessary to verify whether the training data used in Stage 1 and Stage 2 were sampled in a way that preserves the correspondence among brain, heart, and gut data. If this information is already included in the paper, it would be helpful to clarify it more explicitly.

  • Reviewer confidence

    Not confident (1)

  • [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

    I like the idea from this paper where the authors think about using multiple scans of PET, which include many organs (brain, heart, and gut) for early AD diagnosis. The authors propose 3-stage method, which I found useful as well starting from: (i) pretraining the model on brain and whole-body PET images based on self-supervised learning, (ii) feature extraction and alighment for brain-heart-gut model, and (iii) multi-constraint knowledge distilolation for brain-only model from the stage 2 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.

    I think that this paper tackles an underexplored but promising direction by incorporating this brain, heart, and gut information from PET scans for early diagnosis of AD. I believe this multi-organ approach is biologically grounded and reflects recent findings on the systemic nature of AD, adding novelty and clinical relevance. I also found that these three-stage framework is very well-structured and thoughtfully designed. With such a framework then, I think it can has a clinical applicabililty through brain-only inference. Moreover, the paper leverages a multi-cohort dataset combining large numbers of unlabeled and labeled brain and whole-body PET scans. I think this can enhance the robustness of the learned representations.

  • 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.
    • I was initially excited by the paper’s title and abstract, which suggest the integration of heart and gut PET scans, beyond brain scans, for early Alzheimer’s disease diagnosis. However, upon reviewing the methodology, it becomes clear that these multi-organ scans are used solely during training for knowledge distillation, while the final inference model operates on brain-only PET data. While this strategy can be pragmatic given the limited availability of full-body scans, it would be more impactful if multi-organ data were also leveraged at inference time, especially in settings where such scans are available.
    • It is great with the multi-cohort dataset that is used in this paper. However, with such datasets that comes from different sources, there might be unexpected shape or scanner-specific confounders, and a standard pipeline for preprocessing used in the paper might not be enough to avoid these biases [1].
    • In the methods section, it seems like the authors augment the datasets first before splitting, such procedures might lead to data leakage that will lead to overoptimistic results. Best practices recommend splitting first, then applying augmentation only within the training set. This issue has been well-documented in previous studies [2, 3].

    Reference:

    1. Dauchelle, V.W., Grenier, T. & Sdika, M.. (2024). An unexpected confounder: how brain shape can be used to classify MRI scans ?. Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 250:338-351 Available from https://proceedings.mlr.press/v250/dauchelle24a.html.
    2. Yagis, E., Atnafu, S.W., García Seco de Herrera, A., Marzi, C., Scheda, R., Giannelli, M., Tessa, C., Citi, L., Diciotti, S.: Effect of data leakage in brain MRI classification using 2D convolutional neural networks. Scientific Reports 11(1), 22544 (Nov 2021). https://doi.org/10.1038/s41598-021-01681-w
    3. Rumala, D.J. (2023). How You Split Matters: Data Leakage and Subject Characteristics Studies in Longitudinal Brain MRI Analysis. In: Wesarg, S., et al. Clinical Image-Based Procedures, Fairness of AI in Medical Imaging, and Ethical and Philosophical Issues in Medical Imaging. CLIP EPIMI FAIMI 2023 2023 2023. Lecture Notes in Computer Science, vol 14242. Springer, Cham. https://doi.org/10.1007/978-3-031-45249-9_23
  • 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.

  • 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
    1. Please refer to the comments regarding weakness. Can authors give some clarifications on tehse?
    2. Did the authors split or augment the dataset first? I might misunderstand about this step, because the given explanation in the paper is not clear enough for me.
    3. How would the authors tackle the problems related to shape and/or specific-scanner biases due to the multi-cohort datasets?
  • 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 — could be accepted, dependent on rebuttal

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

    I really find the paper to be novel and conceptually impactful, especially with the idea of leveraging PET scans of multiple organs (brain, heart, and gut) for enhancing brain-only models via multi-contraint knowledge distillation for early AD diagnosis. However, the paper might have a critical methodological flaw that weakens the trustworthiness of the reported results.

    Specifically, it is unclear from the paper whether the data augmentation was applied before the dataset splitting, but if so, this practice can introduce data leakage. Such things are well-know in medical imaging and can result in significantly over-optimistic results. This issue really limits the paper’s reliability, even if the proposed framework is well-motivated and novel.

    Moreover, the paper might benefit from deeper discussion relating to the confounding variables (related to shape and scanner-specific biases) across the multi-cohort datasets

  • 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




Author Feedback

We thank all the reviewers for their valuable comments. We have addressed all the reviewers’ comments as follows.

  1. Regarding the feasibility of whole-Body PET acquisition (R1): Thank you for your concern. We confirm that routine acquisition of whole-body PET images is feasible in clinical practice. This is supported by systems like Siemens’ Biograph mCT Flow, United Imaging’s uExplorer, and GE’s Discovery MI, which are already in use in hospitals.
  2. Regarding the different settings at inference time (R1, & R3): We recognize importance of addressing whole-body PET images and potential missing organ data during inference. In Stage II, our brain-heart-gut model fully supports scenarios where whole-body PET are available for testing. However, in clinical practice, AD diagnosis is primarily relying on brain imaging, which is why we focus on brain-only PET in Stage III. To further address the missing data concern, we plan to incorporate additional experiments handling missing data in future research.
  3. Regarding the clinically relevant of the model (R1): Thank you for your comments. Our proposed brain-heart-gut guided brain-only model achieved a 4.9% improvement in AUC. While this improvement is notable, we acknowledge the need to evaluate its clinical relevance. To this end, we will include statistical tests, validate the model on independent external datasets, and conduct a reader study to further validate the effectiveness.
  4. Regarding the clarification of details (R1): We sincerely apologize for the lack of clarity in our manuscript. 1) standard deviations in tables: The reported standard deviations are derived from 5-fold cross-validation. The relatively large values may be attributed to the limited size of the current dataset. However, it is noted that the standard deviations of our method are consistently smaller than those of the comparison methods. We will expand the dataset in future work. 2) PET/MR data acquisition: For public datasets, PET and MRI images were acquired separately, while for in-house datasets, they were collected simultaneously using a PET/MR scanner.
  5. Regarding the case-level correspondence (R2): Thank you for raising this important concern. In Stage I, which is the pretraining phase, the brain, heart, and gut encoders are trained separately, so the data are not from the same individuals. However, in Stage II, we finetune the model on whole-body PET images, the input brain, heart, and gut are from same individual cases to maintaining case-level correspondence. This clarification will be explicitly included in the formal version of the manuscript.
  6. Regarding the biases of multi-cohort dataset (R3): Thank you for your comment and we acknowledge the potential biases from multi-cohort dataset. We agree that addressing these biases is crucial. In our paper, we follow a widely adopted preprocessing pipeline which has been successfully employed in studies such as Maheux et al. (Nature Communications), Chadebec et al. (TPAMI), and Wen et al. (Medical Image Analysis). Additionally, we found the paper you mentioned to be highly insightful and will explore incorporating its methodology into our future work.
  7. Regarding the data augmentation (R3): We sincerely apologize for the confusion in our manuscript. We split the dataset first and apply data augmentation only within the training dataset. This will be clarified in the formal version of the manuscript.
  8. Regarding the mistakes of paper writing (R1): Thank you for your thorough review. We apologize for any occasional spelling and mistakes in the manuscript. We will address these issues and strive for improvements in the final version of the paper.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top