List of Papers Browse by Subject Areas Author List
Abstract
Alzheimer’s disease (AD) is characterized by abnormal amyloid-β (Aβ) deposition, which causes neural damage and cognitive decline. Aβ positron emission tomography (PET) serves as the gold standard for preclinical diagnosis of AD. However, practical limitations, including high costs, radiation exposure, and constrained accessibility, have motivated recent studies to indirectly predict Aβ deposition patterns from MRI data. Unfortunately, existing methods have not fully leveraged the coupled pathological information from both functional and structural brain networks. To address this gap, we propose Graph Reconstruction Aware Fusion (GRAF), a novel framework designed to predict regional Aβ-PET patterns by integrating functional and structural pathological information. GRAF employs a graph-masked autoencoder to learn integrated network topology embeddings by reconstructing masked edges from both functional and structural networks, effectively utilizing node and edge features. Subsequently, the well-trained encoders are fine-tuned to predict regional Aβ patterns. Extensive experimental results demonstrate that our proposed GRAF framework outperforms six state-of-the-art methods. Our code and representative case examples are publicly available at https://github.com/ninicassiel/GRAF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4304_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ninicassiel/GRAF
Link to the Dataset(s)
Huashan dataset : https://www.huashan.org.cn/pet/
BibTex
@InProceedings{YuaHao_AβPET_MICCAI2025,
author = { Yuan, Haoyue AND Liu, Yuxiao AND Liu, Feihong AND Shen, Dinggang},
title = { { Aβ-PET Pattern Prediction via Graph Reconstruction-Aware Fusion (GRAF) of Functional and Structural Networks } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {76 -- 86}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces GRAF (Graph Reconstruction-Aware Fusion), a two-stage self-supervised learning framework that integrates functional and structural brain networks to predict regional Aβ-PET deposition patterns, a critical marker for AD. Its design lies in the node-edge bidirectional encoder and cross-attention mechanism, which allow for fusion of multimodal brain connectivity data while reconstructing masked graph edges as a pretext task.
- 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.
GRAF leverages both edge and node features bidirectionally during encoding. It interestingly integrates node-to-edge and edge-to-node message passing, incorporating topological and microstructural features from brain networks. It also uses a dual-path cross-attention mechanism to refine embeddings from FCNs and SCNs reciprocally, improving representational capacity over simpler fusion strategies like concatenation.
- 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.
-
Only one private dataset is used, with no external validation on public datasets like ADNI or OASIS, which are standard benchmarks in the AD field. This severely limits reproducibility and generalizability.
-
The dataset is imbalanced, and the paper provides no discussion of strategies (e.g., reweighting, balanced sampling) to address this during training.
-
There is no indication of proper hyperparameter tuning across all methods. Using a single set of parameters per method may introduce bias, potentially favoring the proposed method while underestimating baselines.
-
Some of the “state-of-the-art” methods used for comparison (e.g., GCN, GAT) are basic graph models, not truly SOTA for this application domain. More competitive methods from recent AD-related graph learning literature are not considered.
-
No discussion on whether the reported MSE (e.g., 0.0211), MAE, or MAPE are clinically meaningful. Without reference to thresholds accepted in clinical practice or PET quantification literature, it’s unclear if this level of error is acceptable.
-
Figure 2 is too small and cluttered, making it difficult to visually interpret the spatial distribution and accuracy of predicted Aβ patterns.
-
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Despite its technicality in graph-based fusion and network design, the paper suffers from multiple critical limitations that preclude acceptance:
-
Reproducibility is fundamentally compromised by the use of a private dataset, lack of code/data release, and missing details on hyperparameter tuning. The field requires transparency, especially in medical AI.
-
The evaluation is limited and potentially biased, with only one dataset (not publicly available), imbalanced labels, and weak baseline comparisons, leading to questionable generalizability.
-
There is no clinical benchmark to contextualize the model’s prediction error, weakening its claim of feasibility for AD diagnosis support.
-
- 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
Review #2
- Please describe the contribution of the paper
The paper introduces a graph reconstruction-aware fusion (GRAF) method to predict Abeta-PET regional pattern. It uses fused features of both structural connectivity revealed by diffusion MRI and function connectivity revealed by functional MRI.
The network has two main steps: 1. self-supervised network: It encodes both structural and functional connectivity and uses cross-attention to get embedding. The decoder plays the role of reconstructing the input mask. 2. after step 1, the model freeze the encoder and use regressor to prediction the abeta-PET pattern.
The paper compares their method with 6 other methods both qualitatively and quantitatively. It also has a ablation study to evaluate the main components’ capability. The writing is overall very clear.
- 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 paper introduces a novel network architecture to extract the feature of both structural and functional MRI using self-supervised learning method. It compares proposed method with a number of other literatures and demonstrates a significant improvement in the clinical dataset. Equation is demonstrated clearly and results echo the method.
- 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.
An important question is why MRI data would contain information about PET signals. As far as I understand, abeta accelerate aggregation of tau, and tau leads to neurodegeneration, which is reflected in MRI as structural volume loss. This is a one way interaction. However, the neurodegeneration signal is much subtler than the abeta-PET signal, raising the question: does MRI have sufficient information to accurately estimate the PET signal?
In addition, MRI data contains normal aging and disease-related atrophy. From my understanding of the paper, the system is not able to address this important issue.
In Fig. 2, the authors show qualitative comparison. How are the patients selected among hundreds of patients? It lacks a clear demonstration. What are the models’ complexity participated in the comparison? It will be better to provide a brief introduction.
- 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
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?
The paper proposes a novel network architecture to learn self-supervised functional and structural feature. The feature is then used to predict the abeta-PET pattern. The experiment provides good details in comparison with other methods.
However, there is one important question: does the dMRI and fMRI have sufficient information to predict complex PET data? Author should answer this question clearly in the paper. Also, the model can not distinguish the disease caused atrophy and aging related atrophy from MRI. It is also a main drawback of the method.
- 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 address points I proposed. It is an interesting method worth a wider audience.
Review #3
- Please describe the contribution of the paper
This paper introduces GRAF, a two-stage self-supervised learning framework designed to predict regional Aβ-PET patterns by fusing functional connectivity networks (FCNs) and structural connectivity networks (SCNs). The methodology addresses the high cost and limited accessibility of PET imaging for Alzheimer’s Disease (AD) diagnosis. It leverages masked edge reconstruction and cross-attention mechanisms for effective multi-modal representation learning.
- 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 Node-Edge Bidirectional Encoder is a technically interesting approach to jointly leverage node and edge-level information, motivated by the anatomical distinction between gray and white matter.
-
The fusion of FCNs and SCNs is well-executed; the ablation study demonstrates that it outperforms simpler alternatives.
-
The method improves over state-of-the-art (SOTA) baselines in both quantitative metrics (MSE, MAE) and qualitative assessments (visual error maps).
-
- 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 caption of Figure 1 should be improved to clearly describe each subcomponent — especially panel (c), which depicts the custom encoder.
-
In Figure 2, the prediction error for the Aβ-positive subject is approximately double that of the Aβ-negative subject. It may be helpful to stratify the evaluation by Aβ status (e.g., in Table 1), or consider reporting prediction error across different disease stages in future work.
-
There is an inconsistency in Section 2.3: the paper states that the encoders are “frozen” in Stage 2, yet also mentions they are “fine-tuned.” This contradiction requires clarification.
-
- 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
The paper is well-written and clearly presents a method with strong clinical relevance. Additional analysis of model robustness across demographics or disease stages would further strengthen its applicability.
The necessity of the Node-Edge Bidirectional Encoder is not clearly isolated in the ablation study. While the approach is biologically plausible, future work could benefit from comparing it to simpler alternatives, such as standard GCN or GAT encoders, to better assess its standalone contribution.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a methodologically novel and clinically relevant approach with strong empirical results. While there are minor weaknesses and areas for clarification, they do not detract from the overall quality or significance 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.
Accept
- [Post rebuttal] Please justify your final decision from above.
The ideas are promising, and with stronger comparative analysis and better handling of clinical variability the approach could be of interest for the community.
Author Feedback
We appreciate the reviewers’ recognition of our technical novelty and practical value. Below, we concisely address the reviewers’ main concerns:
R2: Rationality of MRI predicting ROI-based PET SUVR: SCN and FCN changes significantly correlate with Aβ pathology, providing effective markers even for preclinical AD. Aβ oligomers disrupt white matter integrity, affecting morphology and network topology [10.1038/s41392-023-01484-7], with initial changes observed in default mode network (DMN) and attention networks [10.1148/radiol.2021210383]. Additionally, abnormal functional connectivity in DMN hubs is a known Aβ-related marker [10.1038/s41582-021-00529-1]. These network features effectively represent early pathological changes, enabling PET prediction via MRI data. Our model explicitly leverages these correlations through node-edge feature integration.
R1/2/3: Clarification of MSE performance and dataset limitations: Fig.2 presents ROI-based errors complementing the whole-brain average in Table 1. Page limits prevented extensive ROI-level reporting. Our findings confirm the feasibility of ROI-based SUVR prediction using SCN-FCN features, reflected by the overall superior MSE across whole brain and 5-fold cross-validations. Nevertheless, we recognize our method’s preliminary limitation, including insufficient handling of demographic variability, disease stages, and data imbalance (fewer Aβ+ subjects), leading to relatively higher errors in the Aβ+ group. Ongoing work addresses these issues through advanced multi-stage pretraining and multitask learning strategies. Additionally, our extended journal version will include public dataset validations (e.g., ADNI).
R1/2/3 Presentation and interpretation in Fig.2: Subjects in Fig.2 were age- and MMSE-matched across Aβ- and Aβ+ groups, revealing notably higher errors in the inferior occipital gyrus in Aβ+ subjects, illustrated by a wider range color bar. Addressing this imbalance remains a priority in our ongoing research. The final paper will enhance visualization clarity and provide detailed ROI-level error analyses.
R2/R3: Dataset selection, baselines, and hyperparameter tuning. To convince reviewers, we provided our codes in the anonymous repository as mentioned in our abstract. Our high-quality in-house dataset (fMRI: voxel size 2³ mm³, timeseries: 488; HARDI: voxel size 1.5³ mm³, two shells, b=1500/3000 s/mm², 90 directions) provides more functionial and structural information than that of ADNI and OASIS. Future validations on ADNI and OASIS datasets are planned to confirm method generalizability. Baseline methods, GCN and GAT, were selected to specifically benchmark our dual-graph fusion approach, alongside three dual-graph fusion SOTA methods. Hyperparameter tuning included optimizer selection (ADMM vs. SGD), learning rates (1e-3 to 1e-5, equispaced across scales: 10 trials in total), combined with weight decay (1e-4 to 1e-6, equispaced across scales: 10 trials in total), and layers per block.
R1: Other minor issues: We appreciate suggestions for manuscript improvements and will revise the all mentioned issues in figure presentation, details involving technical implemention and clarity for reporting our results.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The reviewers note the interesting methodology proposed. clinical relevance, and promising experimental results. However, there were also concerns noted regarding the feasibility of the MRI to PET prediction framework, limited comparisons to baselines, dataset limitations, and missing implementation details/code for reproducibility. Many of these concerns were resolved in rebuttal, including availability of code, rationale for the prediction framework, and explanations regarding the dataset. While more comparisons would certainly strengthen the paper, the work I believe as is would still be of interest to the miccai community. The authors should please ensure that requested clarifications and details are included in the final version of the paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A