List of Papers Browse by Subject Areas Author List
Abstract
Functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) are key tools for characterizing brain function, yet fMRI’s high cost and limited portability hinder large-scale screening and longitudinal monitoring, motivating EEG-based inference of fMRI-derived representations. Existing EEG-to-fMRI methods predominantly perform signal-level regression and are sensitive to hemodynamic variability and EEG spatial ambiguity. Therefore, we reformulate EEG-to-fMRI translation as graph-to-graph mapping in functional connectivity (FC) space to directly target network-level organization with a time-aggregated representation. To fully exploit both scarce paired recordings and abundant unpaired data, we propose a two-stage semi-supervised framework: (i) within-modality FC reconstruction to learn robust modality-specific manifolds, and (ii) paired EEG-to-fMRI translation with asymmetric Teacher-Student latent alignment (fMRI teacher, EEG student), which suppresses modality-specific noise while preserving shared functional structure. We further incorporate hierarchical-order fMRI FC derived from connectivity-profile similarities to capture higher-level network organization. Experiments on simultaneous EEG–fMRI datasets show consistent improvement over strong baselines in FC similarity and network-pattern agreement, and the learned representations further improve downstream Alzheimer’s disease diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0924_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiXia_Bridging_MICCAI2026,
author = { Li, Xiaoye AND Li, Weiqi AND Mi, Jia AND Huang, Shijie AND He, Yuanzhe AND Lin, Xin AND Cui, Zhiming AND Shen, Dinggang},
title = { { Bridging EEG to fMRI: Semi-Supervised Functional Connectivity Translation via Latent Alignment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work reframes EEG-to-fMRI translation as a graph-to-graph mapping problem in functional connectivity space, rather than predicting raw fMRI signals. The authors first train modality-specific encoders on large unpaired EEG and fMRI collections (Stage I), then use a teacher–student latent alignment scheme on the smaller set of paired EEG–fMRI recordings to perform cross-modal FC translation (Stage II). They also test whether the resulting EEG encoder transfers to Alzheimer’s disease classification.
- 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 motivation is laid out clearly: the authors give a reasonable account of why signal-level EEG-to-fMRI regression is fragile (HRF variability, ROI-wise independence) and why FC-space translation could sidestep those issues. The method follows logically from this setup — FC construction feeds into Stage I self-reconstruction, which then initialises Stage II cross-modal alignment, and the paper reads as a coherent pipeline rather than a loose collection of modules. On the numbers side, Table 1 shows gains over every included baseline, and the ablation in Table 2 confirms that each of the three main components (Stage I pretraining, latent alignment, hierarchical-order supervision) contributes positively. I also appreciate that the authors went beyond reconstruction metrics and ran a downstream AD classification experiment (Table 3, frozen encoder + MLP), which at least makes the practical intent of the representation concrete.
- 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 comparison set is too narrow for the claims made. The introduction itself discusses [13] and [14] as prior EEG-to-fMRI methods, yet neither appears in Table 1 — only NeuroBOLT [15] plus four generic graph models. Calling this “strong baselines” in the abstract is hard to justify when the two most directly relevant prior methods are missing.
2.The downstream AD experiment is the weakest link in the evidence chain. Seven EEG datasets are pooled into a single 217 AD / 146 NC split, but several of the cited sources raise questions: PearlNeuro [6] is a database of middle-aged at-risk individuals, not a clean AD-vs-NC benchmark; AD-Auditory [12] involves auditory entrainment rather than standard resting-state; and the APAVA reference [23] is a methods paper on EEG FC topology, not a dataset descriptor. Without knowing how labels were assigned and whether the 5-fold split controls for source identity, the Table 3 improvements could partly reflect dataset or label confounding rather than genuine representational quality.
3.The paired data used in Stage II (115 subjects, 831 sessions aggregated from six sources) is not described at a level that allows external verification. In particular, [5] reads as a research paper on NODDI microstructure, not a standard data release, so it is unclear exactly which sessions it contributes and how they were obtained.
4.Baseline fairness is also only partial. The proposed method starts from source-localised multi-band EEG FC and benefits from large-scale unpaired pretraining, while NeuroBOLT operates at the signal level and its FC is computed post-hoc. Table 1 therefore shows that this particular pipeline works well, but the comparison is not apples-to-apples enough to support broad superiority claims.
5.The novelty is more in the combination than in the individual pieces. The hierarchical-order FC construction follows [32], the source localisation and PEC pipeline draws heavily on [10], and the teacher–student alignment idea is well established. What is new is putting these together for EEG-to-fMRI FC translation — a valid contribution, but the paper sometimes overstates it.
- 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
The paper would benefit from tighter experimental accounting and more careful comparative language. A useful rebuttal would address four points: explain why [13] and [14] were left out of Table 1 (or add them); give a transparent breakdown of how the Stage II paired sessions were aggregated across sources; spell out the label construction and sample selection for PearlNeuro, AD-Auditory, and APAVA in the AD experiment; and draw a clearer line between what is borrowed from [10]/[32] and what the authors consider their own 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.
(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?
I lean toward rejection, though not strongly. The idea of working in FC space rather than at the signal level is sensible, the pipeline hangs together, and the numbers in Tables 1–2 are encouraging within the reported benchmark. What holds me back is that the benchmark itself is incomplete (missing [13] and [14]), the Stage II data aggregation is not transparent enough to audit, and the AD classification experiment pools heterogeneous sources without adequate explanation of labels or confound control. Together these gaps leave the main comparative claim and the clinical-relevance claim under-supported for a venue like MICCAI.
- 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 clarify several points that were previously unclear: the hierarchical operator Ψ now has an explicit formula, Stage II data are broken down by source, and the novelty claim relative to [10] and [32] is toned down. These changes help with readability and go some way toward addressing reproducibility. That said, three evidence quality problems persist. The comparison remains narrower than the claims suggest. Omitting [13] and [14] is understandable on practical grounds, but the manuscript still advertises “strong baselines” and “comprehensive evaluation” when two of the three most cited EEG-to-fMRI methods are missing. The Wilcoxon tests added in the rebuttal do not fill this gap because they lack test unit specification, per baseline p-values, and multiple comparison correction. The KL vs cosine ablation I asked for is also absent. The AD validation, meanwhile, is less clear after the rebuttal than before it. The paper reports 217 AD/146 NC subjects (363 total), yet the per-source counts add up to 269 / 192 (461 total, matching the record count). This unexplained mismatch worsens the label/source confounding concern raised in my original review. The 5-fold split has not been shown to control for source identity, and PearlNeuro (0 AD/78 NC), whose source describes it as a cohort at risk of dementia, is hard to justify as a clean normal-control benchmark. Cross paradigm generalization is claimed but not tested. The rebuttal confirms a random subject level split rather than a paradigm stratified or leave paradigm out design, meaning the same paradigms can appear in both train and test folds. The reported metrics therefore say little about robustness to unseen acquisition conditions.
Review #2
- Please describe the contribution of the paper
The paper proposes a two-stage semi-supervised framework for translating EEG functional connectivity (FC) to fMRI FC via graph-to-graph mapping. The key insight is that cross-modal translation is reformulated from signal-level regression (predicting voxel/ROI time series) to network-level mapping in FC space, making the task less sensitive to hemodynamic response variability and EEG spatial ambiguity. Stage I trains modality-specific ChebGCN encoder-decoder pairs on abundant unpaired EEG and fMRI recordings to learn robust within-modality FC representations. Stage II fine-tunes the EEG encoder on scarce simultaneous EEG-fMRI pairs using asymmetric Teacher-Student latent alignment, where a frozen pretrained fMRI encoder acts as teacher and guides the EEG encoder (student) via instance-level cosine similarity and distribution-level KL divergence losses over graph-level embeddings. A hierarchical-order fMRI FC representation, encoding pairwise correlations between ROI connectivity profiles, is introduced as a richer supervision target to capture higher-order network structure beyond first-order pairwise correlations. The learned EEG representations are validated both on FC translation quality and on downstream Alzheimer’s disease classification using seven external EEG datasets.
- 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 shift from signal-level to FC-space translation is the paper’s clearest conceptual contribution and is well-argued: hemodynamic response variability and temporal-scale mismatch destabilise point-wise regression, whereas FC aggregates over extended windows and captures inter-regional structure directly. 2.The asymmetry between available data modalities is a fundamental constraint in this domain. Exploiting tens of thousands of unpaired recordings for Stage I pretraining while reserving scarce simultaneous EEG-fMRI pairs for cross-modal alignment only is the right inductive bias, and the two-stage separation is clean and principled. 3.The comparison covers both signal-level SOTA (NeuroBOLT) and four graph-level methods of increasing modeling capacity, all evaluated under identical subject-independent splits and the same FC extraction pipeline, enabling fair comparison across paradigms. 4.Freezing the EEG encoder and training a lightweight MLP for Alzheimer’s disease classification on seven independent EEG datasets is a well-designed transfer test.
- 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 hierarchical FC operator Ψ is critically under-specified. The paper describes high-order FC as correlation between pairs of ROI connectivity profiles and associated FC as correlation between first-order and high-order profiles, but provides no formal definition, no equation, and no illustration. The resulting three-channel hierarchical representation A_hier is central to Stage I supervision and Stage II decoding, yet a reader cannot reconstruct it from the text alone. Citing reference [32] is insufficient; the paper must be self-contained. This is a significant reproducibility concern. 2.No statistical significance testing is provided. All FC translation results (Table 1, Table 2) are from a single 7:1:2 split on 115 subjects. Given the high variance (PR std ~0.12), the reported improvements over baselines may not be statistically significant. Paired tests (e.g., Wilcoxon signed-rank on per-subject metrics) or confidence intervals over multiple random splits are necessary to support the claim of consistent improvement. 3.The 115-subject paired dataset is aggregated from six protocols with substantially different cognitive states (resting state, auditory oddball, visual oddball, sleep, eyes open/closed). FC profiles are known to differ systematically across task conditions, and the paper describes the Stage II split only as subject-level 7:1:2 without stating whether it is paradigm-stratified. If subjects from the same paradigm appear in both train and test, the reported FC translation metrics may reflect within-paradigm generalization rather than true cross-subject transfer. The paper does not report performance broken down by paradigm, making it difficult to assess how robust the method is across acquisition conditions. This should be discussed, and paradigm-stratified evaluation or at least per-paradigm reporting would substantially strengthen the experimental section. 4.The KL divergence over embeddings is non-standard. Equation 3 applies softmax to 128-dimensional graph-level embeddings and computes KL divergence between the resulting distributions, treating embedding dimensions as categories with no clear probabilistic interpretation. The rationale for this formulation over alternatives (MMD, cosine loss alone, contrastive loss) is not given, and no ablation isolates its contribution from the cosine loss in Eq. 2.
- 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 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 paper addresses a clinically relevant problem and the core reformulation of EEG-to-fMRI translation in FC space rather than signal space is well-motivated and supported by clear empirical gains over strong baselines. However, three issues are serious enough to recommend rejection in the current form. First, the hierarchical FC operator Ψ, which is central to both training stages, is not formally defined and cannot be reproduced from the text. Second, results are reported on a single small split without significance testing, making the claimed consistent improvement difficult to verify. Third, the projector and the non-standard KL loss over embeddings are both unexplained and unablated. The paper has the substance of an accept but requires revision to meet the reproducibility and experimental rigour standards.
- 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 rebuttal partially addresses my concerns. The two most critical issues, the formal definition of the hierarchical FC operator Ψ and the absence of statistical significance testing, are now resolved. The authors provide explicit equations for H and S, confirming that Ψ is deterministic and reproducible. Paired Wilcoxon tests are reported with p<0.001 for PR and p<0.05 for MSE, which, while not in the original manuscript, establish that the observed improvements are statistically significant
Review #3
- Please describe the contribution of the paper
This paper proposes a method to predict fMRI-like functional connectivity from EEG. Instead of predicting fMRI time signals, the authors translate EEG functional connectivity into fMRI functional connectivity. The method has two stages: first, it learns EEG and fMRI representations separately using unpaired data; then, it uses paired EEG-fMRI data to align EEG features with fMRI features. The paper also uses higher-order fMRI connectivity as extra supervision.
- 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 problem is important because EEG is much cheaper and easier to collect than fMRI. There is mobile EEG nowadays. Predicting fMRI-like network information from EEG could be useful for clinical studies and screening. It is a good idea to work at the FC instead of the raw signal. EEG and fMRI have very different temporal resolution, so direct signal prediction is difficult. FC-level translation seems more stable and more meaningful. The semi-supervised setup is reasonable because paired EEG-fMRI data are limited. Using large unpaired EEG and fMRI datasets before training on paired data is a good idea. The proposed method performs better than the benchmarked methods, and the ablation study supports the main parts of the model.
- 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 Transformer module is not tailored for the symmetric FC matrices. FC matrices have the upper and lower triangles contain duplicate information. If the model uses the full N×N matrix, it may waste memory and computation by processing the same edges twice. This needs clarification. Second, the method is limited to use static FC, so the high temporal resolution of the EEG not benefited. This makes the impact of the method limited. Third, the EEG FC construction is confusing. The paper mentions slow envelope fluctuations below 1 Hz, but also uses five EEG frequency bands. This should be stated clearly. Fourth, the paper does not report EEG channel counts, montages, sampling rates, or how these differ across datasets. This is important because the method relies on EEG source localization before computing ROI-level FC. Fifth, the hierarchical-order FC operator Ψ is not clearly defined. The authors should provide exact equations and clearly state which FC matrices are included.
- 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.
(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?
The problem is important and the main idea is promising, concurrent EEG-fMRI studies are getting popular. Predicting fMRI-like functional connectivity from EEG could be useful, especially since EEG is cheaper and more portable than fMRI. I also like the choice of working at the FC level instead of raw signals, and the semi-supervised setup makes sense because paired EEG-fMRI data are limited. The results are encouraging, with better performance than the baselines and ablations supporting the main components. However, several details are unclear. The Transformer does not seem tailored for symmetric FC matrices, static FC ignores EEG’s temporal advantage, EEG channel/montage information is missing, the slow-envelope/frequency-band explanation is confusing, and the hierarchical FC operator is not clearly defined. Overall, the paper is promising but needs clearer methodology and discussion of limitations.
- 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 rebuttal addresses several of my concerns, related to the hierarchical FC operator and clarifications of the EEG dataset. My concerns related to the static FC and use of full symettric FC matrix remain partially addressed, yet these are more a limitation of the approach rather than a flaw. So, I lean Accept.
Author Feedback
We thank the reviewers for their feedback. Below, we concisely address the reviewers’ concerns: Q1.Comparison fairness & baselines, R1.We agree that [13] and [14] should be discussed more explicitly. We selected NeuroBOLT [15] as the most comparable signal-level EEG-to-fMRI baseline because [13] requires strict EEG-fMRI alignment flags/events, unavailable in parts of our data, making reproduction unreliable; [14] and [15] are from the same group; [15] is newer, more robust, and better suited for resting-state comparison, while [14] has no released code. [15] uses raw signals as intended, so the comparison evaluates signal-level vs. FC-level paradigms under the same FC metric. Q2.Dataset transparency & clarification, R1/R2.Given limited paired subjects and heterogeneous paradigms, we used a random subject-level split across all paired data instead of paradigm-specific training, which makes the task more challenging and indirectly supports cross-paradigm generalization. Stage II data are simultaneous EEG-fMRI recordings, totaling 115 subjects/831 sessions: NATVIEW 22/392, Oddball 17/102, NODDI 15/15, rest & sleep 33/255, Eyes Open/Closed 8/8, and Simultaneous EEG-fMRI 20/59.One session is one simultaneous run. The split is subject-level: train/val/test = 80/12/23 subjects and 582/83/166 sessions. For AD evaluation, it is supportive evidence, not definitive clinical validation. We only selected AD or NC subjects; other diseases or ambiguous diagnoses were excluded. Metadata contain 363 subjects from 461 EEG records, with 217 AD/146 NC subjects. AD/NC counts are ADFSU 160/12, PearlNeuro 0/78, ADFTD 36/29, ADSZ 24/24, BrainLat_EEG 20/28, AD-Auditory 17/10, and APAVA 12/11.The 5-fold split is subject-level. We will add source-wise tables, label rules, corrected citations, and discuss paradigm/source heterogeneity. Q3.Hierarchical operator Ψ, R2/R4.Although Ψ is described conceptually, explicit equations are needed. Given first-order fMRI FC C∈R^(N×N), high-order FC is H_ij=corr(C_(i,:),C_(j,:)), measuring similarity between ROI-wise connectivity profiles. Associated FC is S_ij=corr(C_(i,:),H_(j,:)), capturing cross-order similarity. The target is A_hier=Ψ(C)=[C,H,S]. Ψ is deterministic with no learned parameters. Q4.Statistics & KL, R2.Statistical evidence was insufficiently reported. We had conducted paired Wilcoxon tests on existing held-out predictions but omitted them due to space. Despite large PR variance, our method significantly outperforms all baselines in PR (p<0.001) and MSE (p<0.05). For KL, we will clarify that it is not a decisive latent-alignment loss, but an auxiliary teacher-student distillation regularizer. It softly encourages the EEG student embedding to follow the fMRI teacher’s latent distribution, while cosine similarity and reconstruction remain the main constraints. Q5.Prior components vs. contribution, R1.We follow [10] for standard EEG/fMRI preprocessing. Inspired by [32], we construct hierarchical fMRI FC targets to capture richer connectivity patterns. Our contribution is integrating these into semi-supervised EEG-to-fMRI FC translation: FC-space graph-to-graph mapping, unpaired representation learning, and frozen fMRI-teacher-guided EEG alignment. Q6.Other clarifications, R4.For symmetric FC, full N×N is used not to model duplicated edges independently, but to keep each ROI’s row-wise connectivity profile as a node feature for graph encoding. Upper-triangular inputs are more efficient but lose this ROI-wise structure. EEG FC captures temporal dependencies over a window, though not moment-to-moment dynamics. Most EEG datasets use a 19-channel 10–20 montage, with a few exceptions; after source localization, all data are mapped to the same ROI space, so acquisition differences do not change the model input. Finally, δ/θ/α/β/γ are carrier bands, while “<1 Hz slow envelope fluctuation” refers to envelope dynamics used for PEC.
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.
Overall, the reviewers appreciated the novelty of treating EEG to fMRI translation as a graph matching problem, noting that signal-level inference is not robust. They also appreciated the teacher-student alignment approach and downstream AD classification task. However, they raised several concerns about the paper clarity, method reproducibility, and fairness of the experimental comparisons. The authors should clearly address these and the other reviewer concerns in their rebuttal.
- 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 authors have done a good job addressing the main critiques. This work would be of interest to MICCAI.
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 authors have addressed several concerns, the claims on comprehensive evaluations should be further clarified.
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 is a borderline paper in my track. One reviewer still maintained a negative recommendation and raised several concerns. I partly agree with these concerns, especially regarding the need for clearer experimental accounting and stronger validation. However, some issues were reasonably addressed in the rebuttal, for example the authors clarified that cross-validation was performed at the subject level. Although several aspects remain insufficiently clear, I do not think these concerns fundamentally undermine the overall contribution. Given the relevance of the problem and the potential value of the proposed FC-level EEG-to-fMRI translation framework, I recommend acceptance at this stage.