List of Papers Browse by Subject Areas Author List
Abstract
Functional magnetic resonance imaging (fMRI) provides high-resolution whole-brain activity measurements but is constrained by high cost, limited accessibility, and restrictive acquisition conditions. Recent advances in cross-modal synthesis have enabled the inference of fMRI signals from electroencephalography (EEG), facilitating estimation of whole-brain activity from portable, low-cost recordings in naturalistic settings. Despite promising results, existing approaches rely on Transformer-based encoders that lack inductive biases for the causal, delayed nature of hemodynamic responses, and provide limited insight into how region-specific predictions are formed. In this paper, we propose MARBLE (Mamba-Attention for ROI-conditioned BOLD Estimation), an end-to-end framework for whole-brain EEG-to-fMRI translation. MARBLE employs a hybrid Mamba-Transformer encoder, where selective state-space modeling provides an inductive bias for HRF-related temporal structure while maintaining linear-time complexity. ROI-conditioned decoding heads operating on a shared latent representation generate region-specific BOLD time series. Our approach provides a favorable trade-off between reconstruction performance and computational efficiency, using 3.2x fewer model parameters and 2.3x less peak memory. Reconstructed functional connectivity matrices demonstrate preserved temporal dynamics and inter-regional structure, and cross-attention topographies reveal network-specific electrode contributions without explicit spatial supervision.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4437_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/emeelkaa/marble
Link to the Dataset(s)
https://huggingface.co/datasets/NeurdyLab/NeuroBOLT
BibTex
@InProceedings{KimEmi_MARBLE_MICCAI2026,
author = { Kim, Emil AND Gahm, Jin Kyu},
title = { { MARBLE: Lightweight EEG-to-fMRI Translation via Mamba-Attention and ROI-Conditioned Decoding } },
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
The authors present a novel deep learning model that can translate EEG signals into fMRI signals by employing a temporal Mamba plus Transformer encoder and a ROI-wise decoder, while maintaining region specificity and needing fewer parameters than state-of-the-art models (due to the linearity offered by Mamba). In addition, the decoders’s cross-attention mechanism allows the direct mapping of attention weights to the brain, making the model interpretable ad-hoc. The proposed approach performs comparably or better than the state-of-the-art models used as baseline and can faithfully reconstruct functional connectivity matrices.
- 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 architecture proposed here is novel in the incorporation of temporal Mamba and ROI-wise cross-attention, which reduces the number of parameters of the model and makes projecting attention weights to the brain possible, respectively. 2.The proposed model performs better than state-of-the-art models that it was compared with across most of the metrics presented here.
- 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.There are implementation details missing (data split and sizes of model inputs and outputs). 2.The authors do not discuss any limitations of the model. 3.FC can have negative values too. Authors should either include negative correlations in to the FC matrix, or justify the use of connections that are positive only. 4.Authors do not report statistical significance of their proposed model being better than the state-of-the-art approaches. 5.The authors do not describe how they adapt the LaBraM model (EEG-specific foundation model) to perform the EEG-to-fMRI task, which undermines the validity of using it as a baseline. The authors should better motivate its use. 6.The sample size is limited to 22 subjects, which warrants a generalizability study of the proposed model to out of distribution data.
- 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
Major concerns:
1.Statistical significance of the performance differences between MARBLE and the baselines is not reported. Given the small sample size of 22 subjects, it is not clear whether the improvements shown in Table 1 are meaningful or within the range of expected variability. 2.The functional connectivity matrices shown only include positive correlations. FC matrices can have negative values, and restricting to positive connections only either discards meaningful anti-correlations or implies a specific analytical choice that is not justified anywhere in the manuscript. 3.The sample size of 22 subjects is limited. The manuscript would benefit from a generalizability analysis, or at least a discussion of how the model is expected to perform on out-of-distribution data. 4.The use of LaBraM as a baseline is not well motivated. LaBraM is an EEG-specific foundation model pre-trained for EEG classification tasks, and it is not clear how the authors adapted it to perform the EEG-to-fMRI translation task. The other baselines are sufficient for comparison, but the authors should at minimum clarify how LaBraM was used.
Minor comments:
1.The data split is not described with enough clarity. While a 3:1:1 subject-specific split is mentioned, it is not clear how many subjects fall into each set or how many time points each split contains. 2.The input and output sizes of the model are not explicitly stated. It appears that the model takes 16-second EEG windows as input and predicts one fMRI time point at a time, but this should be stated clearly. 3.The model was trained for only 20 epochs. The authors should justify that this is sufficient for convergence, or report training and validation loss curves to support this. 4.Limitations are not discussed.
- 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 architecture proposed here is genuinely interesting: the combination of temporal Mamba and ROI-wise cross-attention is a principled design choice that addresses known limitations of purely Transformer-based EEG encoders, reduces the parameter count, and enables direct projection of attention weights to the brain. The reported performance is also competitive with, and in most cases better than, the baselines.
However, the evaluation has enough gaps to tip the balance toward a weak reject. The most important concern is the absence of significance testing on a 22-subject dataset, which makes it difficult to assess whether the reported improvements over the state of the art are meaningful or within the range of expected variability. The treatment of functional connectivity is also incomplete, as restricting to positive-only connections without justification leaves an important part of the analysis in question.
None of these weaknesses are fundamental. Adding significance tests and justifying or correcting the FC analysis would substantially strengthen the manuscript. With these revisions, this work would be above the bar for acceptance.
- 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 addressed all four of my major concerns and my four minor clarifications in their response. On the positive-only functional connectivity matrices (Q2), the authors clarified that FC-Corr is computed on the full signed upper-triangular correlation matrix and that the 25% top-edge threshold is applied only to the rightmost panels of Fig. 4 for visualization, not to any quantitative metric. On reread this is consistent with Sec. 3.2, and my original concern reflected a misreading of the figure rather than a substantive analytical issue. On the LaBraM baseline (Q4), the authors clarified that LaBraM is included as a foundation-model-only control, initialized from the LaBraM-base checkpoint with the classification head replaced by the same multi-ROI regression head used across baselines and trained under identical conditions. This is a principled use of LaBraM as a control for the value of EEG-to-fMRI specific design, and resolves my concern. On statistical significance (Q1), the authors did not provide formal significance tests, but argued that mean and standard deviation across 6 held-out scans under a subject-independent protocol consistent with prior work captures the dominant source of variability, and tempered the “superior performance” framing to a “balanced trade-off”. On sample size and generalization (Q3), the authors acknowledged the 22-subject, single-site, eyes-closed cohort as a limitation and committed to stating it explicitly in Sec. 4.The minor clarifications were addressed correctly, with sufficient implementation detail to support reproduction.
In line with the AC guidance that meta-reviewers should not penalize papers for failing to provide new experimental results in the rebuttal, I do not weigh the absence of formal significance testing or an expanded cohort against the paper at this stage. Two of my four major concerns turned out to be misreadings of the manuscript that the rebuttal correctly clarified, and the remaining two are acknowledged as valid limitations that the authors will discuss explicitly in the camera-ready. The novelty of the temporal Mamba plus ROI-conditioned cross-attention design, the competitive performance against state-of-the-art baselines, and the interpretability afforded by ROI-attention topographies outweigh the remaining rigor concerns. I upgrade my recommendation from Weak Reject to Accept.
Review #2
- Please describe the contribution of the paper
This paper proposes MARBLE, a lightweight whole-brain EEG-to-fMRI translation framework. The method combines a Mamba-Transformer encoder with an ROI-conditioned cross-attention decoder to predict ROI-level BOLD time series more efficiently while improving reconstruction performance and preserving functional connectivity structure.
- 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 method improves synthesis performance while reducing the number of parameters and memory usage, which enhances its practical value. 2.The evaluation simultaneously demonstrates signal reconstruction quality and functional connectivity preservation. 3.The ablation study highlights the effectiveness of two key modules, namely the ROI-CA decoder and tMamba. 4.The paper provides an attention-based interpretability analysis, revealing corresponding relationships between brain functional networks and EEG electrodes.
- 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 evidence supporting the main innovation claims is insufficient. The authors argue that tMamba is better suited than a Transformer for modeling HRF-driven delays and unidirectional dependencies, but they do not provide direct evidence for this claim, such as further analysis of whether the temporal dependencies learned by Mamba are consistent with the physiological characteristics of the hemodynamic response. 2.Some results do not support a blanket claim that the proposed method outperforms the baselines. For example, FC-Corr is slightly lower than that of NeuroBOLT and UnEBOLT, so the use of the term “superior performance” in the main text is not fully appropriate. 3.Some implementation details are still missing, such as the specific structures of the encoder and decoder.
- 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 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 acceptance because this paper proposes a well-motivated and practical approach that achieves a good balance among performance, efficiency, and interpretability. Although some claims should be stated more cautiously and some implementation details need clarification, the overall technical contribution and experimental evidence are sufficient for acceptance.
- 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’ response has addressed my concerns well, particularly through the additional implementation details and the more carefully calibrated conclusions based on the experimental results. These revisions will significantly improve the quality of the paper, and therefore I support its acceptance.
Review #3
- Please describe the contribution of the paper
The paper proposes a lightweight architecture for cross-modal synthesis of fMRI signals from EEG. Specifically, a Mamba-based transformer encoder with state-space modeling to capture HRF dynamics and an ROI-conditioned decoder for generating region-specific BOLD time series.
- 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 proposed solution—with the choice of architecture, along domain-specific enhancements (encoder with state-space model and decoder with ROI-conditioning)—provides improvement in performance as a reduction in model size. The model also includes interpretability as an inherent feature.
- 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.
Several conceptual and methodological aspects need more clarification: (a) Resting state fMRI corresponds to spontaneous BOLD fluctuations that are slow, continuous oscillations that reflect intrinsic network activity. So, the use of state-space model for capturing HRF-driven temporal dynamics seems like a conceptual mismatch. (b) EEG has poor sensitivity to deep subcortical structures (hippocampus, thalamus, basal ganglia). Performance comparison of the ROI-conditioned heads on subcortical vs. cortical ROIs would have given more credit to “whole-brain” activity estimation from EEG. (c) Several works have already proposed combining Mamba with Transformers. So major contribuitions are ROI-conditioned decoding heads and the cross-modal EEG-to-fMRI translation objective specifically. (d) An ablation study to observe the contribution of Transformer blocks in addition to the tMamba would have been more insightful. (e) Cross-attention topographies revealing network-specific electrode contributions in Figure 5 would have been more meaning when compared with topographies generated without ROI-CA module.
- 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?
Lack of major novelty factor as well as clinical utility of the proposed method.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and the Area Chair for their constructive feedback. We are encouraged that the reviewers recognize the practical value, efficiency, and interpretability of MARBLE, and we address the main concerns below. (1) Statistical significance (R1-1), superiority claims (R2-2). Table 1 reports mean ± std across 6 held-out scans under a subject-independent protocol consistent with prior studies [16,18]. The standard deviation reflects inter-subject variability, the main source of variation in this setting. MARBLE showed consistent improvements across metrics and Yeo networks (Fig. 2), indicating stable rather than isolated gains. Ablations further support component-specific contributions (ROI-CA: FC-Corr −0.061, tMamba: Corr −0.010). While formal statistical testing could further strengthen the evaluation, we believe the reported trends are meaningful within the benchmark setting used in prior work. Regarding claims, we agree that “superior performance” is too strong given that FC-Corr is slightly lower than [16,18]. We instead claim a balanced trade-off between reconstruction quality, efficiency, and interpretability. (2) Positive-only FC matrices (R1-2). FC was computed using Pearson correlation without thresholding or sign filtering. FC values are predominantly positive in our data. As stated in Sec. 3.2, FC-Corr is computed on the full signed correlation matrix (upper-triangular); the 25% top-edge threshold is applied only to the rightmost panels of Fig. 4 for visualization and does not affect any quantitative metric. (3) Generalization concerns (R1-3). We acknowledge the 22-subject cohort is limited, constraining generalization. The subject-independent protocol follows standard practice [16,18], and we used cross-subject splits to evaluate generalization within available data. We will state the single-site n=22 cohort and eyes-closed-only condition as explicit limitations in Sec. 4.(4) HRF interpretation (R2-1, R3-a), novelty (R3-c), additional ablations (R3-d, R3-e). tMamba was not intended to model hemodynamic responses, but to capture temporally extended EEG dependencies from 16-second EEG windows for regression to slow-varying BOLD signals. The contribution of tMamba is supported by ablation. We will soften “captures HRF dynamics” to “provides inductive bias for HRF-relevant structure” in the revision. The Mamba–Transformer encoder is a deliberate efficiency choice for this domain-specific application. Our central novelty is ROI-CA, which is the largest contributor in our ablation, enabling ad-hoc interpretability via attention topographies (Fig. 5), absent in adapter-based decoders. ROI-specific topography (Fig. 5) is intrinsically defined by the ROI-CA mechanism: without ROI queries, no ROI-conditioned attention map exists to compare against. (5) Missing implementation details (R1, R2). MARBLE uses embedding dim 64, depth 4, Mamba state size 16, conv width 4, dropout 0.2.The ROI-CA employs 64 ROI queries with the same embedding dim. The 3:1:1 split is over subjects, with no subject in multiple splits. Each input is a 16-s, 26-channel EEG window at 200 Hz; the target is the BOLD signal across 64 ROIs at the corresponding fMRI time point. Training for 20 epochs was selected based on monitored validation loss, which plateaued well before termination. (6) LaBraM motivation (R1-4). LaBraM [13] is included as a foundation-model-only control. We initialize from LaBraM-base and replace its classification head with the same multi-ROI regression head used across baselines, under identical training conditions. This isolates the contribution of EEG-to-fMRI specific design over a generic large-scale EEG foundation model. (7) Subcortical sensitivity (R3-b). We acknowledge EEG limitations in subcortical regions. The DiFuMo atlas includes cortical and subcortical ROIs, and results are reported across all regions. A separate cortical-versus-subcortical evaluation is an interesting direction for future work.
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.
The reviews are mixed, with two reviewers leaning toward weak accept and one toward weak reject. The reviewers highlight critical evaluation gaps that undermine the claims from the authors: the lack of statistical significance testing on the small 22-subject dataset, the unjustified restriction to only positive functional connectivity values, and insufficient evidence that the Mamba component genuinely captures hemodynamic response dynamics. Additional concerns include missing implementation details, an unmotivated and unclear LaBraM baseline, limited novelty due to prior Mamba-Transformer hybrids, and the need for analysis of subcortical ROI performance and model limitations. The authors should prepare the rebuttal to address those concerns.
- 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.
In the rebuttal, the authors clarified that the functional connectivity analysis already used signed correlations and that the LaBraM baseline was appropriately adapted as a foundation-model control, resolving two major concerns raised by Reviewer 1.The remaining issues such as small sample size, softened superiority claims, and missing implementation specifics were either addressed through textual clarifications or acknowledged as limitations for the camera-ready version. Two reviewers confirmed their recommendations to Accept after the rebuttal. Based on the final reviewer assessments and the authors’ responsive revisions, the paper can be recommended for 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.
Despite some lingering concerns, the authors have done a good job addressing the main critiques to extent allowed by MICCAI (i.e., no new experiments). This work would be of interest to the community.
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.
The draft proposes a lightweight EEG-to-fMRI translation framework combining a Mamba-Transformer encoder with an ROI-conditioned cross-attention decoder.
The rebuttal provided missing implementation details requested by R1 and R2 (embedding dimensions, ROI query count, training split, input/output dimensions), substantially improving reproducibility. Consequently, R2 upgraded to accept post-rebuttal. However, R1 maintain the weak reject score. Despite this hiccup, I think the paper’s core contributions - extreme parameter efficiency and inherent interpretability via ROI-conditioned cross-attention - are well evidenced.
