Abstract

Resting-state functional MRI (rs-fMRI) provides sensitive network-level biomarkers for a broad range of brain disorders, yet its cost, duration, and limited availability mean that most patients undergo only structural MRI (sMRI), leaving a structural–functional gap in routine practice. We propose CONNECT-4, a hyperedge-based multimodal fusion framework that generates subject-specific 4D rs-fMRI volumes directly from T1-weighted sMRI. CONNECT-4 provides three main contributions: (1) a 4D structural to functional generation pipeline that synthesizes full spatiotemporal BOLD sequences from a single structural scan; (2) a hyperedge multimodal fusion module that constructs a unified hypergraph over anatomy, ROI (region of interest) masks, and connectivity priors to generate temporally coherent BOLD dynamics; and (3) a biologically informed multi-scale loss that jointly aligns voxel-wise signals, ROI-level functional connectivity, and global network properties. CONNECT-4 improves MSE by 57.1%, SSIM by 14.6%, and ROI correlation by 17.8% compared to baselines. When synthetic rs-fMRI is added to sMRI in downstream clinical tasks for dementia classification (5 datasets, 12,653 scans) and epilepsy detection, F1 scores improve by 3.3% and 2.2%, respectively, compared to sMRI-only training. These results suggest hyperedge-based multimodal fusion bridges routine imaging and functional diagnostics. Code is available on GitHub.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1731_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/BioMedIA-MBZUAI/CONNECT-4

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HasSal_CONNECT4_MICCAI2026,
        author = { Hassan, Salma AND Salem, Mostafa AND Elsayed, Ayman AND Oullier, Olivier AND Yaqub, Mohammad},
        title = { { CONNECT-4: Brain Connectivity-Guided Hyperedge Graph Fusion for Structural MRI to 4D Rest Functional MRI Synthesis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The author introduce a framework that formulates rsfMRI synthesis as a systems-level prediction problem, learning subject-specific spatiotemporal patterns conditioned directly on anatomy.

  • 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 addresses a clinically well-defined and practically significant gap: synthesizing 4D rs-fMRI directly from T1-weighted sMRI, a task that goes beyond prior work limited to connectivity matrix prediction or static 3D surrogates. The integration of anatomical ROI information as structured, biologically contextualized text embeddings — incorporating normative volume statistics conditioned on age, sex, and scanner — is a novel and well-motivated design choice that enriches the structural conditioning signal beyond conventional voxel-level representations. The authors further conduct rich experiments to evaluate their model, validating synthesis quality across multiple complementary metrics and demonstrating clinical utility on two downstream tasks spanning 12,653 scans across five independent datasets.

  • 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 mask and ROI graphs rely on a population-averaged DWI connectivity matrix from an external dataset, rather than each subject’s own white-matter tractography. Since structural connectivity is known to vary meaningfully across individuals, it is unclear whether the model generates truly subject-specific rs-fMRI or produces outputs that are largely similar across patients. An analysis of inter-subject variability in the synthesized volumes would be needed to clarify this. The biological validity evaluation is limited to mean BOLD spatial correlation and ROI-level functional connectivity. rs-fMRI carries a much richer set of functional properties — including amplitude of low-frequency fluctuations (ALFF), regional homogeneity (ReHo), functional gradients, and dynamic functional connectivity — and agreement across these measures would be necessary to support the biological validity claims made in the paper. Several reported performance gains in downstream tasks are small in magnitude, with confidence intervals that overlap across conditions. Without paired statistical testing (e. g. , paired t-tests across cross-validation folds), it is not possible to determine whether these differences are statistically meaningful. The synthesis model is trained on A4 and ADNI, both of which also appear in the downstream dementia classification evaluation. This overlap between training and evaluation domains may inflate the reported results, and testing on fully independent cohorts would offer a stronger assessment of generalization.

  • 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

    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 biological validity evaluation is limited to mean BOLD spatial correlation and ROI-level functional connectivity. rs-fMRI carries a much richer set of functional properties — including amplitude of low-frequency fluctuations (ALFF), regional homogeneity (ReHo), functional gradients, and dynamic functional connectivity — and agreement across these measures would be necessary to support the biological validity claims made in the paper.

  • 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.

    Thank you for the authors’ response. While I appreciate the effort to address the comments, I remain concerned that several key issues are not fully resolved in the rebuttal. In particular, many responses rely on future revisions, such as adding inter-subject variability analysis, statistical testing, additional biological validity metrics, and expanded ablation studies, rather than providing concrete evidence in the current rebuttal. Overall, I find the work interesting, but the rebuttal does not yet sufficiently strengthen the central claims.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a framework for generating 4D rs-fMRI from T1-weighted MRI using graph-based multimodal fusion, external connectivity priors, a diffusion transformer, and a temporal UNet. The paper reports improved synthesis metrics over several baselines and downstream gains for dementia and epilepsy classification when synthetic rs-fMRI is added to T1w MRI.

  • 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.This paper addresses an important and clinically significant issue, namely the structural-functional gap observed in routine neuroimaging examinations. 2.The overall framework is ambitious and technically interesting, and the reported synthesis and downstream results are encouraging.

  • 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.Although the paper is framed as synthesizing subject-specific rs-fMRI from T1w MRI, the method also uses strong external priors, including segmentation-derived descriptions, known functional connectivity information, and structural connectivity matrices. As a result, it is unclear how much of the generated signal is actually inferred from T1w anatomy itself. 2.While the evaluation primarily reports spatial metrics, it provides limited evidence that meaningful temporal dynamics or functional variability. In addition, the ablation study does not sufficiently isolate the contribution of T1w versus the external priors. 3.The downstream improvements are interesting, but they do not fully resolve this issue, since synthetic rs-fMRI may simply act as a learned auxiliary representation of structural information rather than a faithful reconstruction of functional biomarkers.

  • 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.

    (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?

    This is a paper on an interesting topic, and while the effort to link structural imaging with functional biomarkers is noteworthy, current evidence does not sufficiently distinguish between that derived from structural anatomy and that provided through prior information on external connectivity.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper addresses a clinically motivated but technically challenging problem: synthesizing subject-specific 4D resting-state fMRI from a single 3D T1-weighted structural MRI in order to narrow the structural–functional gap in routine neuroimaging. To this end, the authors propose CONNECT-4, a framework consisting of three major parts: first, multi-stream representations derived from T1 patches, segmentation-mask-derived ROI descriptions, and ROI-level features; second, a hyperedge-based multimodal graph fusion module that connects image patches, mask patches, and ROI nodes through coverage-weighted hyperedges; and third, a diffusion-based 4D generation pipeline with a temporal UNet decoder and a biologically informed composite loss including voxel, SSIM, temporal coherence, perceptual, and FC-preservation terms. The main claimed innovation is to move structural-to-functional prediction beyond connectivity-matrix prediction toward full 4D rs-fMRI synthesis, with hyperedge-based multimodal fusion as the key methodological contribution.

  • 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.

    S1.The paper tackles a clinically meaningful problem. Compared with T1w MRI, rs-fMRI is more expensive, more time-consuming, and less accessible, while functional network information may still provide complementary value in neurodegenerative and epilepsy-related settings. Therefore, the problem formulation itself has clear practical motivation.

    S2.The method design is technically interesting and reasonably novel. Rather than simply concatenating multimodal features, the paper explicitly constructs image, mask, and ROI graphs, then links patch-level anatomy, segmentation-derived context, and connectivity priors through hyperedges before conditioning the generative model. This is a nontrivial design choice in the structural-to-functional synthesis setting.

    S3.The paper goes beyond reporting synthesis quality alone and attempts to evaluate the utility of synthetic rs-fMRI in downstream dementia classification and epilepsy detection across multiple public cohorts. Although the exact split protocol still needs clarification, this broader evaluation perspective is a positive aspect of the work.

  • 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.

    W1.The manuscript does not sufficiently justify why full 4D rs-fMRI synthesis is the right or necessary target. The paper repeatedly frames its main contribution as “moving beyond FC prediction to full 4D rs-fMRI generation,” but it does not clearly establish that full voxelwise 4D synthesis is preferable to directly predicting ROI time courses, functional connectivity matrices, or graph-theoretic descriptors. Given that the downstream usage in the paper still centers on network-level functional information, the current manuscript reads more like a proposal of a more complex task formulation than a convincing demonstration that this formulation is necessary or advantageous.

    W2.The temporal definition of the 4D target is underdescribed, which weakens the credibility of the 4D synthesis claim itself. A 4D fMRI sequence is defined not only spatially but also by temporal length, frame count, and temporal resolution/TR; however, the manuscript does not clearly report the temporal length of the target rs-fMRI volumes, the number of frames per sample, whether all sequences are normalized to a fixed temporal length, whether variable-length generation is supported, or whether the model simply learns one fixed target shape. The concern is not that the paper explicitly claims arbitrary-length generation, but rather that it does not provide enough information to define the boundary conditions of its temporal synthesis task. For a paper centered on 4D rs-fMRI generation, this is a significant omission.

    W3.The method has a pronounced “stacked modules” character, but the ablation evidence is not clean enough to show that this complexity is necessary. On the input side alone, the framework combines a pretrained segmentation model, a brain MRI foundation model, ROI textual descriptions, a clinical text encoder, radiomics, and an ROI-level foundation model; the graph side further adds three graph streams, a hypergraph, graph attention, and Nodes2Token; the generator uses both DiT and a temporal UNet; and the training objective is a multi-term biologically guided loss. Such complexity is not inherently problematic, but the current ablation study does not convincingly demonstrate that these components are all needed. In particular, Table 1 reports only the comparison between the full model and a combined “No HyperGraph + Loss” variant, which removes two core factors simultaneously and does not isolate the actual contribution of the hypergraph itself versus the biological loss. The manuscript mentions additional studies such as removing individual graph streams, replacing the hypergraph with pairwise graph connections, and comparing Nodes2Token against simpler pooling, but these results are not presented in a sufficiently systematic and transparent way.

    W4.Subject-level isolation between pretrained resources and downstream experimental data is not documented, leaving a serious but unresolved leakage/confounding risk. The paper relies on several pretrained components and external priors, including SynthSeg, brain MRI foundation models, a clinical text encoder, an ROI-level anatomical foundation model, and a pretrained fMRI model used for perceptual loss. However, the manuscript does not state whether the pretraining data for these components are subject-wise disjoint from A4, ADNI, OASIS, or other downstream cohorts used in this study. The issue is not that leakage can be definitively asserted from the manuscript, but that the paper does not provide sufficient information to rule it out. In a setting that combines multiple public neuroimaging models and public neuroimaging cohorts, this directly affects the credibility and fairness of the reported results.

    W5.The claim of evaluation on “completely unseen cohorts” is ambiguous relative to the data usage description. The manuscript states that synthesis training uses A4 + ADNI, while downstream dementia evaluation also includes A4, ADNI, ANMerge, OASIS-3, and OASIS-4, yet simultaneously refers to “completely unseen cohorts.” It would be important to clarify whether “unseen” means unseen subjects, unseen visits, unseen scans, or merely unseen folds within downstream cross-validation. Without a more precise split description, it is difficult for the reader to determine whether any subject-level overlap or other information leakage exists across stages.

    W6.The processing pipeline for synthetic rs-fMRI before downstream analysis is insufficiently described. The manuscript clearly states that real rs-fMRI targets underwent co-registration, slice timing correction, motion correction, spatial smoothing, and temporal filtering before training. However, it does not explain whether the generated 4D rs-fMRI volumes undergo additional preprocessing before downstream analysis, whether ROI time series and FC matrices are recomputed from them, or whether the synthetic outputs are fed directly into the downstream GNN classifier. Since one of the core claims of the paper is that the synthetic outputs carry functional value, this “synthesis → post-processing → analysis” chain needs to be described more explicitly.

    W7.The current evidence supports “useful synthetic representations for downstream prediction” more strongly than “faithful reconstruction of 4D BOLD dynamics.” The manuscript reports mean 3D rs-fMRI signal maps and several aggregate metrics, but does not provide richer temporal validation such as temporal statistics, spectral characteristics, dynamic FC behavior, or more direct neuroscience-specific analyses of the generated time series. As a result, the paper is currently more convincing as evidence that the synthetic outputs capture some stable disease-relevant functional correlates than as evidence that full 4D rs-fMRI dynamics are faithfully reconstructed.

    W8.Some baselines are not fully matched to the task, and the related-work positioning could be sharpened. Several baselines come from generic image translation or generic 4D medical image generation rather than specifically from structural-to-functional neuroimaging. At the same time, while the paper cites relevant recent work on conditional 4D fMRI generation and structure-to-function prediction, it does not clearly distinguish methods that are related but not task-matched from methods that would more directly align with its own task definition.

    W9.Important practical deployment details are missing. For a computationally heavy 4D generation framework that is partly motivated by clinical accessibility, the manuscript does not report inference time per case, GPU memory footprint, or the cost implications of the temporal dimension. These details are not decisive by themselves, but they are relevant to the practical value of the method.

  • 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.

    (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?

    This paper addresses a clinically relevant problem and contains a technically interesting design, especially in its hyperedge-based multimodal fusion and its attempt to connect synthesis quality with downstream utility. The paper is also clearly the result of substantial engineering and experimental effort.

    However, I currently lean toward weak reject for three main reasons. First, the manuscript does not yet make a convincing case that full 4D rs-fMRI synthesis is a necessary target relative to more direct functional targets. Second, critical details are missing regarding the temporal definition of the 4D outputs, the downstream handling of synthetic rs-fMRI, and the split protocol across stages, which makes both the task definition and the experimental pipeline incomplete. Third, the use of multiple pretrained models and external priors introduces a serious but unresolved provenance/overlap concern, and the manuscript does not provide enough information to rule out leakage or confounding. Overall, I think the paper has promise, but the current evidence is not yet sufficient for acceptance. A strong rebuttal clarifying data isolation, temporal setup, downstream processing, and cleaner ablation evidence would improve my assessment.

  • 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 thank reviewers for their valuable feedback. We appreciate the recognized strengths, including 1) tackling a clinically significant gap via the first T1-to-4D rs-fMRI synthesis framework (R1,R2,R3), 2) novel hyperedge fusion design (R2,R3) & 3) rich multi-metric evaluation across 12,653 scans/5 cohorts/2 tasks (R1,R3). We address all comments below & will revise accordingly.

R1,R2,R3: Subject specificity/external priors. We intentionally focus on functional inference from T1 only, as DWI/fMRI is typically unavailable due to cost/time. The only fixed prior is a population DWI matrix used solely as hyperedge weights: it scales message aggregation but injects no feature content. All propagated information comes from subject-specific node features (T1, segmentation, patch-wise ROI distribution, radiomics/FM, normative volume deviations). With a per-subject hypergraph and differing node features, shared edge weights cannot collapse outputs to a template; we will add inter-subject variability analysis confirming this. Pretrained components are frozen, never trained on labels & no overlap with evaluation cohorts; we will tabulate provenance.

R1,R2,R3: Biological validity/faithful reconstruction. Our objective is clinical utility, not exhaustive reconstruction of all neurophysiological features; our evaluation is designed accordingly. We evaluated at multiple levels, voxel/ROI/F2F correlation & fMRI FM perceptual loss, consistent with prior structural-to-functional work, where FC and downstream utility are primary validation targets [1]. Critically, synthetic fMRI alone outperforms T1 (Tab.2-3) indicating information beyond structural cues and real↔synthetic models transfer near-symmetrically (Tab.4); both would be unexpected if outputs were purely structural rederivations. We will further clarify this & add ALFF, ReHo & spectral statistics.

R1,R3: Cohort overlap. A4+ADNI are the only public cohorts with paired T1/fMRI for real-vs-synthetic evaluation. Both are multi-center (>67), reflecting generalizability. “Unseen” covers both internal held-out patients & external evaluation cohorts. The test set is patient disjoint (see Implementation Details). We will reformulate results per cohort.

R3: 4D vs FC/ROI prediction. FC/ROI prediction is parcellation-bound & requires retraining per atlas. 4D output is parcellation-agnostic: ROI time courses, static/dynamic FC & voxelwise descriptors are derivable post-hoc. It also enables direct use of pretrained fMRI FM on voxelwise spatiotemporal data, which ROI/FC outputs cannot use.

R1: Statistical testing. We will add paired t-tests/Wilcoxon across folds (T1 vs T1+synthetic) with effect sizes & corrected p-values.

R3: Temporal definition. Targets are fixed 128-frame, TR 3s, 3 mm isotropic harmonized sequences; no variable-length generation claimed. Full preprocessing will be expanded.

R2,R3: Component ablation. We agree finer-grained analysis aids clarity & will expand into a stepwise ablation isolating DWI prior and each component. The “no hypergraph+loss” row highlights our core contributions’ impact in a single setting, tractable given 4D training cost. We prioritized architectural/design ablations; non-competitive results were omitted for space. Full ablations results will be in the code.

R3: Downstream pipeline. Generated volumes lie in the same preprocessed T1-aligned space as targets; no extra preprocessing applied, the chain is synthesis→feature extraction→GNN classifier. We will clarify this.

R3:Baselines. As the first work on T1-to-4D-rs-fMRI synthesis, no task-specific benchmark exists, hence the generic image-translation and 4D medical baselines.

R3,R1:Deployment/reproducibility. Inference time, memory & hardware will be reported. Due to space, we omit detailed dataset descriptions as they are public & well-documented in their citations. Full configs & preprocessing are in the code, released upon acceptance.

[1] Zalesky et al., Network Neuroscience 8.4 (2024)




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.

    This paper proposes a hypergraph learning method to synthesize fMRI data from structural MRI, segmentation data, and structural connectivity. Details on the rationale for incorporating segmental volume and structural connectivity-guided hyperedges are needed, which are the reviewers’ main concern. Quantitative validation of the synthesized fMRIs also needs to be further explained.

  • 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.

    This paper proposes a hypergraph learning method to synthesize fMRI data from structural MRI, segmentation data, and structural connectivity. Reviews agreed with the importance of the work and the novelty of the proposed method. Reviewers’ feedback was mostly related to details on the rationale for incorporating segmental volume and structural connectivity-guided hyperedges, and Quantitative validation of the synthesized fMRIs. The major part of those questions was addressed by the authors’ rebuttal.



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.

    The reviewers acknowledge the clinical significance and technical novelty of the hyperedge-based multimodal fusion for synthesizing 4D rs-fMRI from T1-weighted MRI. The rebuttal effectively clarifies the use of external priors, commits to additional biological validation, and confirms proper data splits. While concerns remain about the extent of temporal dynamics fidelity and the completeness of certain ablations, these do not outweigh the demonstrated downstream utility and the paper’s role as the first framework targeting full T1-to-4D rs-fMRI generation. The paper is therefore suitable for acceptance.



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.

    This paper presents an interesting hyperedge graph fusion framework for synthesizing 4D resting-state functional MRI from structural MRI, addressing a challenging and potentially impactful cross-modality problem. Although the rebuttal does not fully resolve the reviewers’ concerns and relies partly on promises of future revision rather than closing key evidentiary gaps, the work appears to have retained meaningful support after the rebuttal.



back to top