List of Papers Browse by Subject Areas Author List
Abstract
To predict and generate developmental changes in brain functional connectivity (FC), we propose a framework combining a Graph Attention Autoencoder with Brownian Bridge diffusion. The autoencoder encodes FC into low dimensional features that preserve topology, while the diffusion model anchors the source and target domains and leverages age conditioning for stable trajectory modeling. This approach enables direct prediction and generation of the FC of target-age from FC of initial-age. Validated on longitudinal fMRI data from the CCNP cohort (ages 6–18), the framework outperforms existing generative models on key metrics, including mean squared error (MSE), Pearson correlation coefficient (PCC), and structural similarity index (SSIM), while capturing biologically significant developmental brain regions, providing a reliable tool for predictive FC modeling. Code is available at https://github.com/rongyezhang18-eng/BBdiffusion-FC.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0598_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/rongyezhang18-eng/BBdiffusion-FC
Link to the Dataset(s)
Chinese Color Nest Project(CCNP) dataset: https://www.scidb.cn/en/detail?dataSetId=c81f0e90a51b4cfca348ce4da6ca734e
BibTex
@InProceedings{ZhaRon_Age_MICCAI2026,
author = { Zhang, Rongye AND Xu, Shasha AND Liu, Xiaobo AND Bian, Jing AND Cao, Rui AND Wen, Xin},
title = { { Age Conditional Longitudinal Forecasting of Adolescent Functional Connectivity via Brownian Bridge Diffusion Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a framework that combines Graph Attention Autoencoder and Brownian Bridge diffusion for brain functional connectivity prediction and generation. The proposed method generates target-age FC from initial-age FC. The method is evaluated on an adolescent cohort and shows that it captures biologically meaningful developmental brain 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.
1) Using Brownian bridge diffusion for reconstructing the follow-up data from the baseline data is interesting, and the motivation of this model design is straightforward.
- 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 presentation of the paper is very bad. The method section does not provide definitions of many notations (e. g. , h, o(), a, W, N(i) in Eq. 1 and m_t, S in Eq. 5, age_diff, age_ratio), as well as the GAT structure, without any preliminaries, which makes it difficult to understand the equations, loss functions, and the model structures. 2) How did the $k=30$ of top-k connections selected? The paper lacks an ablation study on $k$. Does it affect the generated connectivity patterns under the same experimental conditions? If so, which ROIs were significantly affected and why is that so? 3) Only one dataset and three baseline methods (DDPM, GAN, Flow-based model) were used for comparison, which significantly limits the scope of the experimental validation. 4) In section 3.5, the paper provides a comparison between the raw early-age and raw late-age FC differences vs. the raw early-age and generated late-age FC differences. Table 4 shows that the similarity between the generated late-age and the raw early-age FC is higher than that of the raw late-age and raw early-age FC. These results mean that the generated FC looks similar to the raw early-age FC, but it does not guarantee any “physiological” plausible changes over time. How does the proposed method ensure that the generated FC reflects physiologically plausible temporal progression, rather than merely retaining features of the early-age FC?
- Please rate the clarity and organization of this paper
Poor
- 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 does not provide sufficient information for reproducibility.
- 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
1) How does a non-Markovian sampling strategy improve inference efficiency, and how much efficiency improvements were observed by using it? 2) How were the baseline hyperparameters tuned?
- 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 presentation of the paper leaves substantial room for improvement, and the choice of baseline methods is insufficient for a comprehensive evaluation of the proposed method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal does not sufficiently address my concerns. For example, it is still vague whether the generated late-age FC has “physiologically plausible” temporal progression, as both the paper and rebuttal consistently indirectly claim that the generated data do not simply copy the real FC. Moreover, the experiment still has limitations as only one dataset is used with 3 non-task-specific baselines, and the authors did not guarantee that additional datasets will be added. Although the authors said they will add more baselines, no specific examples are suggested. Therefore, the paper appears insufficient for publication at present.
Review #2
- Please describe the contribution of the paper
The paper addresses age-conditional longitudinal forecasting of adolescent functional connectivity from resting-state fMRI. Its main contribution is a latent-space framework that combines a graph attention autoencoder for ROI-level FC representation learning with a Brownian Bridge diffusion model for endpoint-conditioned generation. By conditioning on early/late FC representations and age-related covariates, the method is designed to better match subject-specific developmental prediction than standard pure-noise generation. The paper further provides empirical validation on the CCNP cohort with comparisons, ablations, and plausibility analyses.
- 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 studies a meaningful and relatively underexplored problem, and its endpoint-conditioned latent-space formulation is well matched to paired longitudinal FC prediction. 2.The empirical study is reasonably comprehensive, with multiple baselines, quantitative metrics, ablations, cross-atlas analysis, and an initial biological plausibility 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.
1.The top-k graph construction used by the GAT encoder would benefit from clearer specification. In particular, it is not clear from the manuscript whether the graph is built per sample, whether symmetrization is applied, or how graph undirectedness/consistency is handled when top-k neighbors are selected at the node level. 2.The Brownian bridge formulation is not fully transparent in the paper. In particular, the role of fθ is described inconsistently across the manuscript (e.g., as a reverse time drift, a gradient term, and a correction term), while Eqs. (4) and (6) appear to suggest a more specific residual. In addition, the derivation of the reverse-time update in Eqs. (9)–(10) is not sufficiently explicit, and the form of λ2(t) does not seem right. 3.Train/val/test splitting and subject-level separation are not specified; risk of data leakage is non-negligible in paired longitudinal designs and must be clarified. 4.In comparative experiments, baselines are restricted to generic generative models; missing strong FC forecasting baselines. 5.No statistical significance testing across methods; reported gains (e.g., PCC 0.63 vs. 0.61) may not be significant.
- 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 does not provide sufficient information for reproducibility.
- 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
In the legend of figure2, the space between “,” and “B” should be added.
- 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 studies an important and relatively underexplored problem and proposes a task-specific longitudinal FC forecasting framework with reasonably broad empirical evaluation. But the role of fθ and the derivation of the reverse-time update remain insufficiently transparent, some implementation/formulation details are under-specified, and the biological/plausibility claims are somewhat stronger than the direct evidence provided.
- 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 authors’ rebuttal effectively clarifies the methodological ambiguities and provides sufficient evidence to address my concerns regarding experimental rigor and potential data leakage.
Review #3
- Please describe the contribution of the paper
The paper’s main contribution is a new age-conditioned longitudinal forecasting framework for adolescent brain functional connectivity (FC). It combines a Graph Attention Autoencoder (GAT-AE) with a Brownian Bridge diffusion model so that, given a subject’s FC at an earlier age, the model can directly generate that subject’s FC at a later target age. The key idea is to do diffusion in a learned low-dimensional graph latent space, while conditioning the generation process on age information. As shown in the architecture diagram on page 3, this lets the model preserve FC topology, avoid the instability of standard noise-to-data diffusion, and model developmental change as a bridge between observed early-age and target late-age brain states.
- 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.
-
Novel problem formulation: age-conditioned FC-to-FC longitudinal forecasting. Instead of doing cross-sectional analysis or generic connectome generation, the paper tackles the harder task of predicting a subject’s future functional connectivity from earlier FC plus target age, which is a meaningful and relatively underexplored formulation for developmental neuroscience. This is interesting because it directly addresses sparse and irregular longitudinal sampling.
-
Interesting technical combination of graph learning and Brownian Bridge diffusion. The method combines a GAT-based autoencoder with Brownian Bridge diffusion in latent space. The novelty is not just using diffusion, but using a bridge process anchored by source and target domains rather than standard pure-noise diffusion. That makes it well suited to paired developmental prediction, where the task is to transform early-age FC into late-age FC in a stable, conditioned way.
-
Well-executed ablation study. The ablations test key design choices: removing the AE, replacing AE with VAE, varying sampling steps, latent dimension, and noise schedule. This is valuable because it shows which components actually matter, especially the importance of the autoencoder and the Brownian-Bridge-specific variance design.
-
Use of real longitudinal developmental data. The study uses the CCNP longitudinal cohort and constructs within-subject scan pairs across ages 6–18.That is a practical strength because truly longitudinal data are harder to obtain and make the task more convincing than purely synthetic or cross-sectional setups.
-
- 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 “Brownian Bridge for longitudinal neuroimaging forecasting” angle is not entirely new. A close prior example already exists in neuroimaging: Stoyanov et al. (2025) proposed a conditional Brownian Bridge diffusion model for cortical-thickness trajectory forecasting. That does not make this paper uninteresting, but it does narrow the novelty claim: the distinctive part here is specifically functional-connectivity forecasting with age conditioning, not Brownian-Bridge-based neuroimaging trajectory prediction as a whole.
-
The train/test protocol is not described clearly enough to rule out optimistic results. The dataset is expanded from 393 scans into 1,380 within-subject scan pairs, but the paper does not clearly state whether the split is performed at the subject level or only at the pair level. If pairs from the same individual appear in both training and test sets, performance could be inflated because the model would effectively see very similar subject-specific patterns during training.
-
Clinical feasibility is not really demonstrated. All experiments are on a single healthy developmental cohort from CCNP, spanning ages 6–18, with no external validation cohort, no neurological or psychiatric patient group, and no downstream clinical endpoint such as diagnosis, prognosis, or treatment response. So while the work is scientifically interesting, any implication of clinical usefulness is still premature.
-
Robustness across parcellations looks limited. The best results are obtained on AAL116, while performance drops noticeably on CC200 and CC400.That suggests the method may be somewhat sensitive to atlas choice and granularity, which weakens the claim that it is a generally robust FC forecasting framework.
-
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend Weak Accept because, despite several important limitations, the paper still makes a meaningful contribution to longitudinal functional-connectivity forecasting. In particular, the task is relevant and the proposed age-conditioned FC forecasting framework appears technically sound and empirically promising. However, I do not rate it higher because the novelty claim is narrower than suggested given prior Brownian-Bridge-based neuroimaging forecasting work, the train/test split protocol is not described clearly enough to exclude possible subject leakage, clinical feasibility is not yet demonstrated beyond a single healthy cohort, and robustness across parcellations appears limited. Overall, the paper is interesting and worthy of acceptance, but these concerns reduce my confidence in its novelty, generalizability, and practical impact.
- Reviewer confidence
Somewhat confident (2)
- [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 is understandable, and there is no change to the initial decision of “weak accept.”
Author Feedback
Regarding Brownian Bridge and GATAE,the notation of fθ may cause confusion.In our paper,fθ is not a diffusion score function or reverse-time drift,but a Brownian Bridge residual estimator.Given the bridge state,time step,and condition,the network predicts the residual between this state and the clean lateage latent endpoint.During sampling,this residual is subtracted from the bridge state to estimate the lateage latent endpoint.We will clarify that m_t=t/T is Brownian Bridge interpolation coefficient.S is a fixed maxvar hyperparameter controlling noise strength.Specifically,the corrected form is lambda_2(t)=m_{t-1}-m_t(delta_{t-1}/delta_t)((1-m_t)/(1-m_{t-1})),which defines the time-dependent coefficient of the earlyage endpoint.The reverse step will be rewritten as an endpoint-estimation based Brownian Bridge sampling step:fθ estimates the lateage endpoint from the current bridge state.The next state is computed using this endpoint,the earlyage endpoint,and a variance-based correction term.agediff/ratio augment age cues.For nonMarkovian sampling,200 steps reduce denoising calls compared with 1K steps while keeping similar performance.
For GATAE,hi,W,a_ij,o,and N_i denote the ROI node feature,projection matrix,attention parameter,activation function,and topk neighbor set.For each FC,the graph is built by selecting topk neighbors from each ROI row.The topk graph is a sample-specific computational graph for GAT message passing,while the FC itself remains symmetric for reconstruction and evaluation.We used k=30 and add k=20/60 ablations:k=20 gives MSE/PCC/SSIM/Cos of 0.29/0.61/0.47/0.96,and k=60 gives 0.31/0.60/0.46/0.95.These results show that k=30 better balances sparsity and performance.
Regarding earlyFC copying,we agree that high similarity to early FC requires clarification.Table4 is used to examine whether the generated change magnitude is close to the true longitudinal change.For AAL,rawearly vs rawlate gives PCC/SSIM about 0.63/0.47,while rawearly vs generated late gives 0.61/0.45,lower than the true longitudinal similarity.Thus, Table 4 does not support the concern raised by Reviewer1.Since age gaps are small and individual FC has strong fingerprints,generated late FC should preserve subject-specific organization while adding age-related changes.If the model copied earlyFC as predicted lateFC,it would be the early-copy baseline,which gives MSE=0.56,PCC=0.63,SSIM=0.46,and Cos=0.93.Our method has lower MSE and higher SSIM/Cos,indicating that it does not simply copy earlyFC.
Regarding data splitting,the concern mainly comes from the insufficient description of the subject level protocol in the original manuscript.We carefully rechecked the implementation to confirm whether the experiments followed subject level separation. To make the protocol fully explicit,we further repeated the experiments under the strictly defined subject level split with multiple random seeds and conducted statistical analyses.The results show that the conclusions from the main comparisons(shared HPs-epochs,etc.for fairness) and ablations remain consistent under this protocol.In the revision, we will clarify the split, report the results, and add paired tests with 95% CIs.
For baselines,early-to-late FC prediction is a conditional generative forecasting task.DDPM,etc. were selected as representative generative baselines,while existing studies mainly focus on cross-modal SC-FC generation and lack a widely accepted reproducible longitudinal FC forecasting baseline.To address this concern,we will further explore and add rare longitudinal FC forecasting baselines.
Finally, we will frame the results as methodological feasibility on a healthy cohort.The cross-atlas analysis was designed to examine how ROI granularity affects prediction and to justify AAL116 as the main atlas.AAL116 best fits the current model and data scale,while CC200/400 remain usable but degrade as more ROIs increase dimensionality,sparsity,noise sensitivity,subject variability.
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.
I recommend an Invite for Rebuttal. The reviewers recognized the importance and relevance of the problem—age-conditioned longitudinal functional connectivity forecasting—and appreciated the technical combination of a Graph Attention Autoencoder with a Brownian Bridge diffusion model. However, there are significant methodological and presentation concerns that must be addressed. A primary issue raised by all reviewers is the lack of clarity regarding the train/validation/test splitting protocol; the authors must explicitly confirm whether the splits were performed at the subject level to completely rule out data leakage between within-subject scan pairs. Methodologically, the reviewers noted insufficient transparency in the Brownian Bridge formulation (specifically the inconsistent role of f_theta, the reverse-time update derivation in Eqs. 9-10, and the form of \lambda_2(t)) and in the GAT structure (e.g., how the top-k graph construction handles undirectedness and consistency). The rebuttal must clarify these mathematical details and define all notations (e.g., h, o(), a, W, \mathcal{N}(i), m_t, S, age_diff, age_ratio). Furthermore, the evaluation needs strengthening: reviewers highlighted the need for a stronger, domain-specific FC forecasting baseline rather than relying solely on generic generative models (DDPM, GAN, Flow-based), and requested an ablation study on the choice of k=30 for the top-k connections. Finally, the authors should temper their clinical claims given the single healthy cohort used, address the performance drop observed across different parcellations (CC200/CC400), and explain how the generated FC ensures physiologically plausible temporal progression rather than simply retaining early-age features.
- 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 paper introduces a novel and well-motivated approach for age-conditioned longitudinal functional connectivity forecasting using a Brownian Bridge diffusion model, which is naturally suited for paired developmental prediction. In their rebuttal, the authors successfully addressed the most critical concerns regarding potential data leakage by verifying their results under a strict subject-level data split, while also clarifying the mathematical formulation of their diffusion process. Although one reviewer remained concerned about the lack of domain-specific baselines, the rigorous re-evaluation and methodological clarifications convinced the majority of reviewers that the work is technically sound and warrants acceptance.
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.
On balance the authors managed to address most of the key reviewer concerns including: subject level data splitting, the formulation of the Brownian Bridge.
The description of the ablation study and comparison methods was not very convincing. However as per MICCAI guidelines it is not appropriate to ask for new experiments in the rebuttal, and on balance these new experiments are not essential to the core claims of the authors.
The key sticking issue is the potential for copying of early FC in the model, which the reviewer that raised this point is not convinced by the response. Indeed the authors should consider adding more explanation in the manuscript and also explain the limitations/assumptions on relying on similarity measures to identify copying of measures.
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.
All reviewers were satistfied with the rebuttal. The authors should address the remaining concerns in the camera-ready version.