List of Papers Browse by Subject Areas Author List
Abstract
Decoding visual content from non-invasive EEG is fundamentally challenged by representation collapse, semantic ambiguity, and stochastic generation drift. In this work, we propose NeuroFlow, a structured EEG–vision alignment framework that addresses these issues from both representation learning and generation perspectives. To capture transient functional connectivity without inducing dimensional collapse, we introduce a Manifold-Aware Dynamic Graph Encoder regularized by an orthogonality constraint. To further resolve the inherent semantic ambiguity of these encoded features, a Prototype-Guided Semantic Rectification layer is employed to anchor loose EEG representations to the compact support of the visual manifold. Furthermore, we reformulate the generation process using Deterministic Flow Matching, constructing an optimal transport path that ensures stable semantic transport by mitigating the stochastic drift of diffusion models. Experiments on the THINGS-EEG dataset demonstrate that NeuroFlow outperforms state-of-the-art methods in zero-shot retrieval (Top-1 Acc: 32.4\%, +4.6\% gain) and achieves superior generation fidelity with 5$\times$ fewer inference steps. Codes are available at https://github.com/happyeverymoment/NeuroFlow
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0221_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/happyeverymoment/NeuroFlow
Link to the Dataset(s)
BibTex
@InProceedings{WanLei_NeuroFlow_MICCAI2026,
author = { Wang, Lei AND Diao, Yueqin AND Yu, Hanyi AND Yang, Yuetong AND Xu, Yanwu},
title = { { NeuroFlow: Manifold-Aware Topological Flow Matching for EEG 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 main contribution of this paper is the proposal of NeuroFlow, a novel framework designed to accurately reconstruct visual images from EEG signals. It tackles three major bottlenecks in existing methods: a) Instead of treating brain channels independently, it uses a dynamic graph (DRG) to capture how different brain regions connect, and a special rule is designed to prevent the extracted features from collapsing into useless noise. b) Since EEG signals are noisy and hard to read, the model uses a “memory bank” of clear visual prototypes. c) Instead of using slow, random diffusion models, it uses Conditional Flow Matching to create a fast, straight-line path from the brain signal directly to the image, making the generation process stable and faster.
- 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 model achieves state-of-the-art accuracy in zero-shot tasks, successfully matching brain signals to the correct images significantly better than existing methods. 2.It drastically speeds up the image generation process, requiring fewer steps than traditional diffusion models while still producing higher-quality and more accurate pictures. 3.The system’s unique design cleverly prevents the common problem of “dimensional collapse” (where data features blur together), making the feature extraction process much more stable and reliable
- 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 model’s performance is currently only validated on a single benchmark dataset (THINGS-EEG) under a specific zero-shot protocol. The authors acknowledge that they have not yet tested the framework across diverse EEG recording conditions or varying equipment, which is critical for proving real-world reliability in Brain-Computer Interface applications. 2.Because the model prioritizes aligning the overall semantic meaning rather than fitting high-frequency noise, the generated images may lack exact fine-grained, pixel-level textures compared to the original visual stimuli.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
It introduces an innovative framework that effectively solves critical bottlenecks in EEG decoding, specifically dimensional collapse and semantic ambiguity. The methodology is well-designed, and the results are highly compelling.
- 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 has adequately addressed my main concerns, so I am satisfied with the authors’ response and recommend acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes NeuroFlow, a two-stage EEG-to-image decoding framework built on: 1) a dynamic graph EEG encoder with Barlow-Twins-style decorrelation, 2) a prototype-guided semantic rectification module, and 3) conditional flow matching for image-embedding generation followed by frozen SDXL + IP-Adapter decoding. On THINGS-EEG, the paper reports improved zero-shot retrieval and competitive generation quality with substantially fewer sampling steps than diffusion-based baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a real bottleneck in EEG-based visual decoding: unstable conditioning and expensive diffusion sampling. 2.The use of flow matching is a meaningful technical contribution in this domain and is a reasonable alternative to diffusion for faster inference. 3.The retrieval gains in Table 1 are nontrivial.
- The paper includes ablation and interpretability analyses rather than only headline results.
- The method is modular and the overall pipeline is easy to follow.
- 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 full evaluation is conducted on only THINGS-EEG (Sec. 3.1). For a MICCAI paper, this is a weak validation base, especially for a method claiming broad robustness to semantic ambiguity and stochastic drift. There is no external dataset, no cross-dataset test, and no serious analysis of robustness across acquisition conditions.
- In Table 2, ATM is shown at 50 steps and NeuroFlow at 10/20/50 steps, but CoCapturer has no inference time reported, and the comparison mixes methods with different sampling regimes and likely different back-end decoders. It is therefore difficult to determine how much of the gain comes from the proposed transport model versus implementation choices around the frozen generative backbone. 3.Table 1 reports mean ± std across subjects, but there are no significance tests, confidence intervals on method differences, or paired comparisons. Given only 10 subjects, the claim of clear superiority should be supported more rigorously. 4.EEG methods are highly sensitive to inter-subject variability. The paper reports aggregate subject-wise performance, but does not deeply examine cross-subject transfer, leave-one-subject-out settings, or calibration sensitivity. This is a major omission for practical relevance. 5.Ablation evidence is still incomplete. Table 3 supports that the full combination helps, but it does not isolate whether gains mainly come from SigLIP, the prototype memory, the BT regularizer, or the dynamic graph itself under matched backbones and parameter counts. In particular, the paper argues strongly for topology modeling, but the ablation also shows the graph module alone hurts. 6.The paper emphasizes better “high-fidelity” generation, but the reported image metrics are mixed, and some are not especially strong evidence of semantic correctness in EEG-conditioned generation. More semantically grounded evaluation would be needed. 7.The work is closer to brain-computer interface / multimodal generative modeling than to core medical image computing. It may still interest MICCAI readers, but the paper should make the medical or neuroimaging relevance more explicit.
- 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.
(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 has a good central idea, and the flow-matching component is the most convincing novelty. But for MICCAI, the current evidence base is too narrow and the evaluation is not yet rigorous enough to fully support the strength of the claims.
- Reviewer confidence
Somewhat confident (2)
- [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
1.A decorrelated dynamic graph encoder that captures transient functional connectivity while mitigating over-smoothing and dimensional collapse.
2.A prototype-guided semantic rectification module that anchors noisy EEG embeddings to visual prototypes for better cross-modal alignment.
3.A flow matching-based generation framework that replaces stochastic diffusion trajectories with deterministic transport, improving generation stability and inference efficiency.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper presents the method as a set of modules, each designed to address a different challenge in EEG visual decoding.
2.The paper goes beyond overall performance and provides extensive ablation and analysis experiments.
3.The proposed method shows consistent performance gains over the compared baselines.
- 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 paper frames representation collapse, semantic ambiguity, and stochastic generation drift as key bottlenecks, but their existence and practical impact in EEG visual decoding are not rigorously established. In addition, many of the cited references are prior visual decoding papers, and it is unclear whether they directly support these claims as the main causes of performance degradation. More direct evidence or more appropriate references would strengthen the motivation.
2.Although each module is intuitively motivated, the overall framework appears largely as a combination of existing components, such as dynamic graph modeling, prototype-based refinement, and flow matching. As a result, the originality of the method and the necessity of each individual component are not fully convincing.
3.While the paper presents flow matching as an important alternative to stochastic diffusion-based generation, in practice it mainly operates as an embedding-space transport module before a frozen SDXL/IP-Adapter backbone. As such, its contribution appears closer to replacing the diffusion prior at the embedding level than to fundamentally changing the overall generative pipeline.
4.The overall flow of the paper is difficult to follow, and the connection between the overview and the method section is not sufficiently clear. In addition, the terminology, notation, and module descriptions are not explained in a sufficiently coherent manner, making the proposed framework feel more like a loose combination of modules than a well-integrated design.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
I encourage the authors to clarify the above concerns in the rebuttal, as my final assessment will depend on how convincingly these issues are addressed.
- 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?
While the paper is reasonably polished and supported by experimental effort, I find its central motivation and technical claims too ambiguous to be fully convincing. The claimed bottlenecks are not rigorously validated in the EEG decoding setting, the method appears largely as a combination of existing components, and the efficiency advantage is not established through a fully fair comparison. The paper is also difficult to follow at times, as the connection between the overview and the method section is not always clear, and the framework does not read as a fully well-integrated design. Overall, the paper is promising, but in its current form it remains marginally below the acceptance threshold.
- Reviewer confidence
Very confident (4)
- [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 provides additional claims and results, but it does not convincingly address my fundamental concerns about the paper’s motivation, originality, and methodological coherence, which would require substantial revision beyond rebuttal-stage clarification.
Author Feedback
We thank all reviewers for acknowledging our retrieval performance, computational efficiency, and interpretability, particularly R1 for validating our methodological novelty. We address the core concerns below:
1.Generation Efficiency & Fair Comparison (R1, R2, R3): We clarify that “high-fidelity” primarily denotes semantic (CLIP, Inception) and structural (SSIM) alignment (our semantic scores are superior, as shown in Tab. 2 and Fig. 2). We acknowledge a trade-off in exact pixel-level similarity (PixCorr), as Flow Matching prioritizes semantic manifold alignment rather than fitting high-frequency stochastic details. For a fair comparison, we re-evaluated CoCapturer using the identical frozen backbone in our revision. NeuroFlow achieves better semantic scores in just 20 steps with a >5x speedup, confirming its intrinsic advantage.
2.Cross-Dataset Validation (R1, R2): Evaluated on an MEG dataset under identical protocols, NeuroFlow (Top-1: 14.8%, Top-5: 37.2%) significantly outperforms SOTA methods NICE (10.8%, 32.1%) and CoCapturer (11.3%, 35.0%), demonstrating robust cross-dataset generalization.
3.Component Synergy & Ablation (R2, R3): Ablation shows strong synergy, not mere combination. DRG alone causes topological over-smoothing and performance drops (Fig. 3). Conversely, BT loss alone yields limited improvement lacking dynamic topological guidance. Only their integration mitigates dimensional collapse and preserves features, boosting performance to 30.2%. SigLIP mainly accelerates training with marginal gains (Top-1: 31.9% to 32.4%), thus omitted due to space.
4.Statistical Significance & Cross-Subject (R2): Paired t-tests across 10 subjects confirm NeuroFlow’s improvements are statistically significant (vs. CoCapturer: Top-1 p=0.02, Top-5 p=0.009). Under the stringent Leave-One-Subject-Out (LOSO) setting, our model achieved 12.9% (Top-1) and 31.4% (Top-5) accuracy, significantly outperforming CoCapturer (9.3%, 25.4%), proving robustness to inter-subject variability.
5.Clinical Value & MICCAI Scope (R2): By decoding neural-visual mappings, our research provides a crucial prognostic evaluation tool for populations with cognitive or visual impairments. This aligns perfectly with MICCAI’s core mission of advancing clinical translation and neuro-engineering.
6.Sensitivity & Robustness (R2): Analyses indicate that performance fluctuations from adjusting key hyperparameters lack statistical significance, showing high parameter robustness. We will explore multi-center data for further validation.
7.Theoretical Motivation & References (R3): EEG “semantic ambiguity” intrinsically originates from skull attenuation and volume conduction, while GNN “representation collapse” is a mathematically proven phenomenon. Specifically, dimensional collapse is evidenced by feature redundancy in Fig. 3(d), and semantic ambiguity by the scattered latent distribution before PRL (Fig. 3(b)). Regarding stochastic drift, deterministic transport yields more stable semantic trajectories under weak EEG conditions. We replaced generic visual decoding citations with specific neuroscience and graph theory literature to strengthen our premise.
8.Methodological Novelty (R3): Our framework is a customized, causally-linked design for EEG, not a loose combination. DRG introduces dynamic topology but causes over-smoothing, prompting the decorrelation (BT) constraint; PRL then corrects the noisy geometry; finally, Flow Matching ensures stable transport to the visual space (showing excellent stability across 16.5k paired samples). We do not aim to replace the entire SDXL framework; our core contribution lies in replacing stochastic diffusion with deterministic transport in the embedding space under weak EEG conditions, significantly enhancing generation stability and reducing inference costs.
9.Clarity of Presentation (R3): We will rewrite the methodology section to unify mathematical notations and logical flow, improving overall readability.
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 reviewers agree that the paper addresses an interesting problem and proposes a promising framework, particularly with the use of flow matching for efficient EEG-to-image decoding. The overall approach is well-structured and shows competitive performance. However, several key concerns remain. The main limitation is the weak empirical validation, as experiments are conducted only on a single dataset without analysis of cross-subject variability or robustness. In addition, the statistical significance of the reported improvements is unclear. There are also concerns regarding the clarity of the technical contribution, as it is not sufficiently demonstrated how much each component contributes to the final performance. Furthermore, the fairness of comparisons (e.g., differences in sampling steps and missing baseline details) needs clarification. Finally, the novelty and motivation should be more clearly justified. The final decision will depend on how convincingly these points are addressed.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a well-structured framework for EEG-to-image reconstruction and addresses several practical challenges in EEG visual decoding. The proposed NeuroFlow combines dynamic graph modeling, prototype-guided semantic rectification, and flow matching in a coherent manner, and demonstrates consistent improvements in both retrieval and generation tasks. In addition, the paper provides extensive ablation studies and analysis experiments that help support the effectiveness of the individual components. The reviewers raised valid concerns regarding validation on a single dataset, limited robustness analysis across acquisition conditions and subjects, and the degree of novelty of individual modules. However, I think these limitations primarily concern the scope of validation rather than the soundness of the proposed methodology. The experimental results are consistently strong, the technical contributions are well motivated, and the analysis goes beyond simple performance reporting. Overall, while additional validation on diverse EEG datasets would further strengthen the work, I believe the paper makes a meaningful contribution to EEG-based visual decoding and provides sufficient technical and empirical evidence to merit 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.
After carefully reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors have provided a thorough, evidence-based response that directly addresses the core concerns raised by all three reviewers.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal provides several clarifications however the concerns related to motivation and novelty still exist.