List of Papers Browse by Subject Areas Author List
Abstract
Assessments for disorders of consciousness are frequently confounded by absent or inconsistent behavioral output, particularly in cognitive motor dissociation, motivating objective and interpretable EEG-based visual-semantic evidence as an adjunct to bedside evaluation. We present EEGRFusion, an EEG-to-image framework that projects EEG into a visual-semantic space and enables (i) zero-shot image retrieval and (ii) EEG-conditioned image generation, together providing clinician-reviewable cues of stimulus-evoked processing. For robust EEG representation, we introduce MAMD, a montage-aware temporal dual-path encoder with a Mamba backbone. For generation, to address the CLIP semantic bottleneck, EEG variability and related issues, we learn an EEG-to-CLIP conditional prior using Rectified Flow and decode it with a frozen diffusion backbone equipped with an IP-Control Fusion Adapter to mitigate semantic and structural drifts. Experiments on THINGS-EEG demonstrate improvements in retrieval accuracy and reconstruction quality over baselines. We further provide uncertainty heatmaps demonstrating multi-sample variability to facilitate risk-aware interpretation. The code is available at https://github.com/hcy1026/EEGRFusion.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4361_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hcy1026/EEGRFusion
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HonChe_EEGRFusion_MICCAI2026,
author = { Hong, Chenyuan AND Diao, Yueqin AND Wang, Lei AND Huang, Ziyan AND Zhao, Jiongning AND Yu, Hanyi AND Xu, Yanwu},
title = { { EEGRFusion: Uncertainty-Aware EEG-to-Image Evidence for Adjunct Bedside Assessment of CMD in Disorders of Consciousness } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces MAMD, a montage-aware temporal dual-path encoder with a Mamba backbone
Experiments on THINGS-EEG dataset
Develops RFFA, a Rectified-Flow-based prior and Fusion Adapter controlled diffusion generator that trains a Control Adapter
- 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.
Develops RFFA, a Rectified-Flow-based prior and Fusion Adapter controlled diffusion generator that trains a Control Adapter.
Introduces MAMD, a montage-aware temporal dual-path encoder with a Mamba backbone
Experiments on THINGS-EEG dataset
- 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 dataset EEG corpus is acquired from 10 subjects, will this be representative of the population.
Could provide some more recent background references
- 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.
(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?
Good layout and detailed diagrams in Fig. 1 and confusion matrix Fig. 2.Interesting topic and contribution.
- Reviewer confidence
Somewhat confident (2)
- [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 have addressed the majority of the review comments.
Review #2
- Please describe the contribution of the paper
This paper presents a method for generating images from EEG signals to support clinical behavioral assessment in patients with Cognitive Motor Dissociation (CMD). The key contributions are: (1) a novel EEG encoder that produces stable EEG representations; (2) a Rectified Flow-based prior that enables sampling of a CLIP image embedding conditioned on an EEG representation; (3) a Control Adapter module appended after the IP-Adapter and SDXL-turbo pipeline to address structural drift inherent in images generated based on an CLIP image embedding, and to produce uncertainty heatmaps that enhance clinical interpretability. Together, these additions improve the fidelity of EEG-conditioned image synthesis.
- 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.
- Clarity and justification of design choices: The paper is well-written and easy to follow. Each design choice is clearly motivated and the training procedure is described in sufficient detail to support reproducibility.
- Novel EEG encoder with strong empirical justification: The proposed EEG encoder is compared against multiple baseline EEG encoders using retrieval accuracy as a generator-independent proxy metric. The results justify the architectural choices for the EEG encoder before the generation pipeline is introduced.
- Well-motivated use of the Control Adapter: The Control Adapter addresses the structural drift inherent in CLIP-conditioned generation. This is a known limitation of CLIP embeddings, which encode high-level semantic information at the expense of low-level spatial and structural information. Moreover, the curriculum training strategy for training the Control Adapter is a clever solution to the cold-start problem: by using blurred ground-truth images as control images and gradually transitioning to SDXL-Turbo outputs, the Control Adapter learns reliable structural conditioning before being exposed to noisier, model-generated control images.
- Limitations are acknowledged.
- 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 clinical motivation requires stronger justification: The model is trained and evaluated on THINGS-EEG2, which consists of healthy subjects viewing naturalistic images under an RSVP paradigm. CMD patients represent a fundamentally different population. The authors should justify why a model trained on healthy-subject data can be expected to transfer to this clinical population.
- The authors should clarify whether the model is trained and tested on data from the same subject (within-subject evaluation) or evaluated on held-out subjects (cross-subject evaluation).
- Tables 3 and 4 (reconstruction quality and ablation study) report results only for Subject 8.The authors should justify why the results are not averaged across all subjects to give a fairer picture of the method’s performance.
- 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?
The proposed method is methodologically interesting, and the paper is clearly presented.
- 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.
My decision has not changed. Additionally, the authors responded to my concerns.
Review #3
- Please describe the contribution of the paper
The reviewer did not understand the paper sufficiently well to point out contributions.
- 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 reviewer did not understand the paper sufficiently well to point out any strength.
- 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 literal presentation of the paper is very poor, and large parts appear to be AI generated: a) The abstract and introduction are particularly poor, and appear to be 100% AI generated. After reading both, the reviewer still has no idea on what is the exact task to be solved in this paper. What is EEG-to-image? What exactly is visualized from the EEG in the image? How does this relate to CLIP? How does Diffusion come into play here? How is any of these related to DoC? What is assessed at “bedside”, and why is this important? b) Some kind of “representation boundaries” are mentioned several times without explaining what this would be. c) It remains mysterious what the “noise endpoint” and the “data endpoint” (section 2.2) should be. Can the authors provide these in their own terms? Is one of these related to images, or to EEG embeddings? d) What would be a “structural drift”, “layout”, “semantics”, “IP-Control” (section 2.3)? At which location do the authors extract “UNet features”, from which kind of data, and what is the dimensionality of such features? What is SDXL-Turbo?
2.The description of the dual-path model, and other parts of the pipeline are insufficient and partially wrong: a) For their temporal path, the authors indicate that the batch dimension (B) is changed. This is clearly wrong. The input needs to be shaped as BxTxC and not TxBxC. b) The authors mention “fourier feature electrode coordinate encoding” and “coarse region priors” without providing any hint on what this is or how they have applied this. c) Since Bi-Mamba blocks build the main blocks in their module, a brief introduction on these blocks would be required. d) It is mentioned that temporal path and spatial path are fused, but not how. There is no final overview of the model topology. e) The authors mention a “CLIP loss”, but it is not clear what the “CLIP loss” is. Do they mean unsupervised contrastive loss? Similarly, it is not clear what the “MSE loss” is, and on which elements it is computed.
3.The experimental setup, and evaluation metrics are unclear: a) The authors mention “training image conditions” and “test image conditions”, without describing what such conditions would be. Again, the task to be solved (RSVP) is unclear. b) Section 3.2 discusses zero-shot identification, but it is unclear what should be identified, and how. c) Other evaluation metrics are just named (2-way z-score, 2-way-acc) without telling what are the two ways. d) The reviewer does not understand how the uncertainty heatmaps in figure 2 are computed, or what they are showing. Why is a pixel-wise uncertainty important, and not a global, image-wise uncertainty?
- Please rate the clarity and organization of this paper
Poor
- 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
-
The license of the “anonymized” source code includes a copyright for Chenyuan Hong, indicating an author and infringing anonymization. The code repeatedly reference directories in /home/diaoyueqin, indicating another author name.
-
Graphics are stored and imported as pixel graphics (.jpg/.png). Better options would be vector graphics (.eps/.pdf).
-
- 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.
(1) Strong Reject — must be rejected due to major flaws
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
For a non-expert, this paper is illegible. Many concepts are assumed to be known. While claimed several times, the application of the method as bedtime assessment tool is unclear. The exact task to be solved remains mysterious throughout the entire paper, and it is unclear what the evaluation metrics evaluate. Anonymization of the code is not done properly.
- Reviewer confidence
Not confident (1)
- [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 provided code clearly failed to pass the anonymization test. Why were the authors invited to write a rebuttal?
Many of my questions were not answered by the authors. I would not trust that the paper improves to a reasonable state where it could be publishable, without additional review.
Author Feedback
We thank all reviewers for your constructive comments. We appreciate the recognition of our “motivated contribution”(R1&2), “methodologically novel design”(R1&2), and “reproducibility”(R1-3). Below we address general concerns followed by specific points. 1.Dataset representativeness and clinical translation(R1-3, AC). THINGS-EEG is a large-scale 82,160 trials and representative benchmark released by the THINGS initiative and leading institutions, and is widely used in EEG-to-image studies (ATM, UBP etc.). Healthy-subject data are not directly deployable to patients, so our work is claimed as algorithmic validation of stimulus-evoked EEG-to-visual-evidence decoding, as a first translation step before patient-specific calibration. Although DoC/CMD EEG introduces differences, healthy EEG also has large inter-subject variability; our results support robustness to variability, making the model promising for clinical transfer. CMD patients may retain stimulus processing without overt behavior, making behavioral assessment insufficient[1-4]. Decoded visual evidence from EEG signals can provide adjunct, clinician-reviewable support. We will add references and perform DoC/CMD validation in the future. 2.Experimental details(R2&R3). Retrieval is within-subject. Following published baselines, we report generation results on Subject 8 for fairness. Image conditions and RSVP are from THINGS-EEG acquisition protocol, not our task. We will clarify these points and further explain zero-shot N-way retrieval identification: N candidates include a true image and N-1 distractors, ranked by EEG-image embedding similarity. The 2-way metric is the N=2 case. Pixel-wise uncertainty is computed from 10 stochastic generations and can reflect both object reliability and structural stability. Unstable/failed generations make the object uncertain; structural drift affects background regions. From another perspective, the heatmap shows that Control Adapter reduces layout instability and thus better helps clinicians identify regions requiring caution.
To R3, due to page limits, some details common in EEG-to-image studies are not fully restated. Gaining more related knowledge from [7] may be helpful for better understanding. 3.Task definition, readability and AI use(AC). Due to page limits, many abbreviations are used, we will remove them except MAMD and RFFA. EEG-to-image is not to visualize raw EEG traces, but to infer visual evidence from stimulus-evoked EEG: retrieval of the viewed image and reconstruction of a stimulus-consistent image. CLIP provides the visual embedding; diffusion synthesizes the image. AI was not used to generate or write the content but only for grammar/language polishing. 4.MAMD details. The batch dimension B is not changed. The permutation only follows PyTorch Transformer’s sequence-first convention. We will clarify notation and expand Fourier electrode encoding, region priors, and Bi-Mamba. Dual-path representations are fused by addition as shown in Fig.1(overview). Our paper defines CLIP loss and MSE loss and states paired EEG-image embeddings. Rectified Flow separately uses MSE between predicted and target velocities. We will add formulas for clarity. 5.Generation details. “Representation boundaries” means the ability of EEG embedding to represent visual information. In Rectified Flow, the noise endpoint is a Gaussian noise vector, the data endpoint is the target CLIP image embedding, EEG embedding as condition. Semantics denotes object/category content; layout denotes edges/background. Structural drift means alteration of original layout. SDXL-Turbo is a pretrained latent diffusion image generator with a denoising U-Net; U-Net features are its multi-scale intermediate maps. “IP-Control” is our module combining IP- and Control-Adapter. 6.Code and figures. I appreciate your note. This is my first submission and I lack experience. I will carefully remove personal paths and license metadata and use vector graphics.
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 paper needs some improvement in writing for readability when there are too many abbreviations used. Also need more justification on how a model trained on healthy subjects will be translated to patient populations.
- 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 concerns raised have been addressed.
Meta-review #2
- 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 paper explores an interesting and clinically motivated direction by investigating EEG-to-image generation as adjunct evidence for bedside assessment in disorders of consciousness. Reviewer 2 appreciated several aspects of the work, including the motivation, the proposed EEG encoder, and the integration of a Rectified-Flow prior with a Control Adapter framework.
However, after carefully reviewing the paper, the reviews, and the rebuttal, I do not believe the current submission reaches the level required for acceptance. In particular, I found Reviewer 3’s concerns regarding the paper’s clarity, methodological presentation, and overall readability to be substantial and difficult to dismiss. A major issue is that the core task formulation, model pipeline, and evaluation methodology are often insufficiently explained, making it difficult. While Reviewer 1 provided a positive recommendation, the review itself was relatively limited in technical depth and did not substantially engage with many of the concerns later raised by the other reviewers. Overall, although the topic is interesting and the proposed direction has potential, I do not recommend acceptance in its current form.
Reviewer 3 noted a potential anonymization issue, citing a copyright notice containing an author name and directory paths revealing another individual’s identity. I was unable to locate these identifiers in the currently available materials and therefore could not independently verify the concern. However, it is possible that such information was present in an earlier version accessible during the review process and later removed. If so, the anonymization concern would be legitimate.
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 reviewers raised concerns regarding the clinical translation of a model trained on healthy-subject EEG data, the clarity of the experimental protocol, and the readability of the manuscript. The rebuttal addressed these concerns satisfactorily by clarifying that the work is intended as an algorithmic validation of EEG-to-image decoding rather than a clinically validated DoC/CMD deployment system. The authors also clarified the evaluation protocol, the use of Subject 8 for consistency with prior work, the interpretation of uncertainty heatmaps, and several methodological details that were insufficiently explained in the original submission.
While the clinical applicability to DoC/CMD patients remains a limitation, the paper appropriately positions this as future translational work rather than a demonstrated clinical claim. The methodological contribution is technically sound, supported by comprehensive experiments, and reproducible through the provided code and implementation details. Overall, I believe the major concerns were sufficiently addressed and the paper is above the acceptance threshold.