List of Papers Browse by Subject Areas Author List
Abstract
EEG-to-fMRI synthesis aims to generate high spatial-temporal resolution fMRI, pivotal for diagnosing brain diseases, using accessible EEG signals. However, current static regression methods fail to account for regional variations in EEG-fMRI time delays and cannot disentangle spatially mixed EEG signals. To address this, we propose DSTP, a framework that explicitly parameterizes this dynamic spatial-temporal mapping. We introduce a Dynamic Heterogeneous Graph Neural Network (DHGNN) to model the evolving mapping between scalp electrodes and brain regions of interest (ROIs). Moreover, we embed a learnable Gaussian Mask into the message-passing mechanism. This acts as an adaptive temporal gate, effectively filtering the EEG history to identify the specific time window that aligns with each ROI’s hemodynamic response. Additionally, a Transformer module integrates the historical evolution of these states, while a multi-objective loss preserves temporal dynamics and brain topological consistency alongside reconstruction accuracy. Extensive experiments on multiple real-world datasets demonstrate that DSTP achieves state-of-the-art performance. Notably, the model exhibits superior zero-shot generalization across tasks and spatial resolutions, proving its ability to learn intrinsic physiological mapping rules rather than overfitting to specific distributions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2443_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{ZheJia_Dynamic_MICCAI2026,
author = { Zheng, Jialun AND Liu, Jie AND Cao, Jiannong AND Saxena, Divya},
title = { { Dynamic Spatial-Temporal Mapping for EEG-to-fMRI Synthesis via Gaussian-Guided Heterogeneous Graph Neural Network } },
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 paper proposes DSTP, a dynamic spatiotemporal mapping framework for EEG-to-fMRI synthesis. The method introduces a dynamic heterogeneous graph neural network to model the time-varying relationships between EEG electrodes and fMRI ROIs. The authors further propose a learnable Gaussian mask to model ROI-specific hemodynamic delays. Transformer-based temporal modules and a multi-objective loss are further used to improve reconstruction fidelity and preserve functional connectivity. Experimental results on resting-state and auditory-task datasets show that the proposed method outperforms several baselines and demonstrates promising cross-task and cross-resolution generalization ability.
- 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 method has a certain degree of novelty. The authors propose using a Gaussian function with mean and variance parameters to model ROI-specific hemodynamic functions, and apply this function as a temporal mask in the message-passing mechanism, reflecting the authors’ unique insights into neural signal analysis and deep learning. -The method description is relatively clear, and the correspondence between formulae and symbols is good. -The experimental design is relatively complete, including comparisons with baseline methods, ablation experiments covering the basic modules, and additional validation on an independent dataset. -The dataset split is handled reasonably, and comparisons are made under both intra-subject and inter-subject settings.
- 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 authors mention two main challenges in the Introduction, among which 2) Spatial Ambiguity is not addressed in a particularly unique way in the paper. -The authors emphasize the subject specificity of the HRF in the cited literature, but the Gaussian mask modeled in the paper seems to have only region specificity. -The HRF usually does not follow a unimodal Gaussian distribution, and a single pair of $\mu$ and $\sigma$ parameters may not fully capture the region specificity of the HRF. Moreover, this modeling choice is not well supported by the experiments. -In Section 2.1, the authors mention that the constructed heterogeneous graph has directed Electrode-to-ROI edges $\epsilon_{ER}$ and undirected Functional edges $\epsilon_{FC}$. The definitions of these two variables should be explained there; otherwise, they may confuse readers. -The authors state that the constructed heterogeneous graph is dynamic, but this dynamic property seems to depend only on the time-varying attention parameters in message passing. Therefore, the dynamic graph structure should be more clearly distinguished from dynamic message passing. -The third term in Equation (5) refers to the neighborhood of node $k$. If FC is not sparsified, this almost means that the neighborhood of $k$ contains all other nodes. The expression of the formula could be simplified here to reduce symbolic redundancy. -Why is the edge weight in Equation (3) obtained by using the fMRI ROI node state at the previous time step $(t-1)$ and the electrode features at the current time step (after hemodynamic Gaussian masking)? What is the physiological basis for this design? How does the formula behave at $t=0$? -Some references do not follow the required format. -The main acquisition parameters for EEG and fMRI are not provided. -Some important information is missing, which reduces the reproducibility of the paper, such as the atlas used for fMRI ROI signal extraction, the time window length $L$, the feature dimension $d$, and the latent dimension. -This citation is inappropriate: Unlike methods relying on static network structure [2], our graph structure dynamically evolves [20, 19, 4, 13], allowing the model to capture time-varying dependencies between the electrodes and ROIs. -There is inappropriate bolding in Table 2: for the intra-subject setting and ConnMSE, DSTP (Ours) is not the best result.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although some issues still remain, the paper may still be suitable for publication at MICCAI as a preliminary study, given its methodological novelty and relatively sufficient experimental design, provided that the above weaknesses are adequately addressed and revised, especially by adding the missing information to improve reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes the DSTP framework to address the core challenges in EEG-to-fMRI synthesis and validates the model’s performance and generalization ability through relatively comprehensive experiments. The authors construct a dynamic heterogeneous graph, using EEG electrodes and fMRI ROIs as graph nodes, and model the spatiotemporal mapping relationship through dynamically evolving edges between the electrodes and ROIs. The paper further uses Gaussian functions to model region-specific hemodynamic response functions (HRFs) and introduces them into the message-passing process of the graph neural network. In addition, the authors design a multi-objective optimization function that includes reconstruction loss, correlation loss, and functional connectivity loss. Experimental results on multiple EEG-fMRI datasets show that DSTP achieves strong performance on temporal correlation, functional connectivity correlation, and other metrics under both intra-subject and inter-subject settings.
- 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 has a certain degree of novelty, especially the combination of a dynamic heterogeneous graph and Gaussian-guided message passing, which is broadly consistent with the underlying physiological mechanism. The learnable Gaussian mask is used to simulate region-specific hemodynamic responses and serves as a weighting factor in message passing, helping the model identify the EEG history segments most relevant to fMRI generation. 2.A multi-objective optimization strategy is adopted, taking into account temporal correlation loss and FC matrix reconstruction loss in addition to reconstruction accuracy. This design is well aligned with the main application scenarios of fMRI. 3.The overall experimental results are strong. From the reported results, DSTP achieves the best overall performance relative to the baselines under both intra-subject and inter-subject settings. 4.The ablation study is relatively complete and covers the effectiveness of the main modules.
- 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 is a lack of direct validation of physiological plausibility. At present, the effectiveness of the Gaussian mask design for simulating ROI-specific HRFs is only validated through fMRI reconstruction performance, without directly examining whether the learned \mu and \sigma are physiologically interpretable. 2.Some experimental results are not fully consistent with the description in the paper. In Table 2, the proposed method does not achieve the best Conn MSE under the intra-subject setting, and this trade-off is not discussed in the manuscript. 3.Some important details are missing. The specific dimensionalities of parameters in the model structure, as well as the time-window length L and feature dimension d, are not reported explicitly, which limits the reproducibility of the paper. 4.The manuscript does not clearly define the reported mean and variance values. It is unclear whether these statistics are computed across time points or across cross-validation results.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the paper still has some issues, its overall level of innovation and experimental support is sufficient for publication as preliminary work at MICCAI.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper addresses the challenges of spatial ambiguity arising from volume conduction and the individual/regional variability of the hemodynamic response function (HRF) in EEG-to-fMRI synthesis. The authors propose DSTP, a dynamic spatial-temporal mapping framework comprising two core components: (1) a Dynamic Heterogeneous Graph Neural Network (DHGNN) that explicitly models the evolving mapping between scalp electrodes and brain regions, and (2) a learnable Gaussian mask-guided message passing mechanism that serves as an adaptive temporal gate to capture ROI-specific HRF delays and dispersions. Experiments on resting-state and task-based datasets are reported, demonstrating performance improvements over existing methods.
- 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.
Strengths and Contributions Physiologically Grounded Problem Formulation: The manuscript correctly frames EEG-fMRI synthesis as a dynamic spatial-temporal mapping problem rather than a static regression task. This perspective directly addresses the core challenges of volume conduction and regional HRF heterogeneity, reflecting sound theoretical insight.
Interpretable and Rational Methodological Design: The learnable Gaussian mask elegantly incorporates HRF priors as a soft constraint, balancing data-driven flexibility with physiological interpretability. The DHGNN, with its heterogeneous nodes and dynamic edges, effectively models the non-stationary cross-modal spatial correspondence. The multi-objective loss function is appropriately designed to preserve functional connectivity topology.
Promising Direction for Generalizability Assessment: The paper presents preliminary evidence of zero-shot generalization across tasks (resting-state to auditory task) and spatial resolutions (DiFuMo 64 to 512), suggesting that the model may capture intrinsic physiological mapping principles rather than merely overfitting to dataset-specific distributions.
- 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.
Lack of Statistical Rigor in Experimental Evaluation: The performance comparisons presented in Tables 1 and 2 entirely omit statistical significance testing (e.g., paired t-test, Wilcoxon signed-rank test). Given the reported standard deviations (e.g., Intra-subject Tcorr: 0.458±0.173 vs. 0.416±0.166), superiority based solely on mean rank ordering cannot be confidently distinguished from random variation. This omission substantially undermines the credibility of the reported empirical gains for a MICCAI submission.
Internal Inconsistency in Methodological Design: The functional connectivity weights in Equation (5) are pre-computed from the training set and remain fixed during both training and inference. This design choice appears at odds with the central premise of “dynamic evolution” and may introduce systematic bias in the presence of domain shifts (e.g., inter-subject evaluation). The authors do not provide sufficient justification for this static prior.
Insufficient Detail in Generalization Experiments: The description of the zero-shot generalization experiments in Section 3.4 is overly cursory. Critical details regarding the source-domain training and target-domain testing protocols, as well as whether baseline methods (e.g., UnEBOLT) received equivalent pre-training and data splits, are absent. This lack of detail prevents independent assessment of the reported findings.
Relatively Limited Sample Size: The primary experiments rely on resting-state data from 22 subjects (29 scans) and task-based data from 10 subjects (16 scans). For a model architecture incorporating Transformer-based temporal modules and dynamic graph structures, this sample size poses a non-trivial risk of overfitting and limits the generalizability of the conclusions.
Incomplete Writing and Compliance Statements: The manuscript lacks an Institutional Review Board (IRB) approval statement concerning data usage, as well as any indication of data availability or code release intent. Furthermore, a dedicated Limitations section is absent.
- 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.
(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?
First, the problem investigated in this manuscript holds clear clinical and neuroscientific significance. EEG-to-fMRI cross-modal synthesis is an important direction in brain imaging research, and the two core challenges addressed—spatial ambiguity due to volume conduction and regional heterogeneity of the hemodynamic response function—are well identified. Framing the task as dynamic spatial-temporal mapping rather than static regression reflects an understanding of the underlying physiology, and the problem formulation is valid.
Second, the methodological design is purposeful and exhibits a degree of novelty. The Dynamic Heterogeneous Graph Neural Network explicitly distinguishes electrode and ROI nodes to model time-varying spatial correspondences, while the learnable Gaussian mask introduces HRF priors as soft constraints, balancing data-driven flexibility with physiological interpretability. The multi-objective loss function, which jointly constrains reconstruction accuracy, temporal correlation, and functional connectivity topology, is thoughtfully constructed.
Third, the experimental results demonstrate performance improvements over existing methods, and the preliminary zero-shot generalization experiments across tasks and spatial resolutions suggest that the model may capture intrinsic physiological mapping principles rather than merely fitting dataset-specific distributions.
Nevertheless, the current manuscript contains a critical shortcoming: the complete absence of statistical significance testing. The performance comparisons presented in Tables 1 and 2 report only means and standard deviations, without any p-values or inferential statistics. Given the substantial standard deviations and modest differences in mean values for several metrics, it is not possible to reliably determine whether the observed gains of DSTP reflect genuine performance advantages or random variation. For a MICCAI submission, this omission materially undermines the credibility of the empirical conclusions.
Additional concerns include a logical inconsistency between the static functional connectivity weights and the central premise of dynamic evolution, insufficient detail in the description of the generalization experiments, a relatively limited sample size, and missing statements regarding ethics approval and limitations.
In summary, this reviewer acknowledges the contributions of the manuscript in terms of problem formulation and methodological design. However, the notable deficiencies in experimental rigor prevent an unconditional recommendation for acceptance in its current form.
- 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 meta-reviewer and reviewers for their constructive feedback and for recognizing the novelty of our dynamic heterogeneous graph framework and Gaussian-guided temporal modeling for EEG-to-fMRI synthesis, as well as the physiologically motivated design and the multi-objective loss. We have carefully considered the major concerns and grouped them into the following categories for clarification Physiological interpretation and scope (R1-C1/R1-C2/R1-C3/R1-C7, R2-C1, Meta-C3). Our goal is not to suggest that a single Gaussian can fully characterize the complete HRF or all subject-specific hemodynamic variability. Instead, the Gaussian mask is intended as a compact ROI-conditioned temporal gating mechanism that highlights the most relevant EEG history during message passing, while broader temporal dependencies are further modeled by the downstream temporal module. We also recognize that the current evidence for the Gaussian mask remains primarily empirical and reconstruction-based, while a more direct physiological interpretation of the learned μ and σ parameters remains an important direction for future investigation. Dynamic graph clarification and methodological consistency (R1-C4/R1-C5/R1-C6/R1-C11, R3-C2). We appreciate the reviewers’ comments on the distinction between dynamic graph construction and dynamic message passing. In the current model, the main time-varying component lies in the Gaussian-guided electrode-to-ROI interaction together with the evolving ROI states, whereas the functional-connectivity term in Eq. (5) is used as a fixed prior estimated from the training data. We recognize that the present wording may leave room for ambiguity, and the role of dynamic cross-modal interaction versus fixed structural prior could be more clearly articulated
Experimental rigor and result interpretation (R1-C12, R2-C2/R2-C4, R3-C1/R3-C3/R3-C4, Meta-C1). We appreciate the reviewers’ thoughtful comments regarding statistical rigor, protocol clarity, and the calibration of the reported conclusions. Regarding Table 2, we recognize that under the intra-subject setting the full model does not achieve the lowest ConnMSE in the ablation study, and this trade-off could have been described more carefully in the manuscript. At the current stage, we view the results as showing the strongest overall empirical trade-off across major metrics and settings, rather than uniform superiority on every ablation entry. We also acknowledge that formal statistical significance testing is not included in the present manuscript, such analysis could provide additional perspective on the reported comparisons, and we appreciate this as a valuable consideration for subsequent refinement. As for the data sample issues, given that paired EEG-fMRI is extremely rare, we have to follow the existing works’ usage of the available dataset with a limited sample size. Reproducibility, reporting, and compliance (R1-C8/R1-C9/R1-C10, R2-C3, R3-C5, Meta-C2). We appreciate the reviewers’ concerns regarding reproducibility and reporting completeness. The manuscript already specifies several parts of the training setup, including the optimizer, training epochs, batch size, learning rate schedule, Gaussian-mask initialization, loss weights, but we recognize that some implementation and dataset details are not stated as explicitly as they could be, including certain acquisition and dimensional settings. We also recognize that the present submission does not include an explicit code-availability or limitations statement, and we are positively considering making implementation details and configuration settings publicly available in the future.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
All three reviewers generally support acceptance (two accept, one weak accept) and recognize the paper’s novel dynamic heterogeneous graph neural network with a learnable Gaussian mask to model region‑specific hemodynamic delays for EEG‑to‑fMRI synthesis. They appreciate the physiologically motivated design and the multi‑objective loss. However, they consistently point to insufficient experimental rigor, including a complete absence of statistical significance tests, missing reproducibility details (model dimensions, time window, atlas), and incomplete validation of the Gaussian mask’s physiological plausibility. It is strongly suggested that the authors thoroughly address the reviewers’ concerns to raise confidence in the reported advantages.
