List of Papers Browse by Subject Areas Author List
Abstract
Understanding visual processing is a foundational pursuit in computational neuroscience. However, effective EEG-visual alignment is hindered by a severe information asymmetry. This disparity manifests as a massive capacity gap between dense image pixels and sparse EEG signals, compounded by a semantic mismatch where image backgrounds overshadow the primary subject, and physiological noise contaminates the neural data. Consequently, existing data-driven alignment methods—lacking explicit structural constraints—often overfit to superficial visual details, thereby limiting zero-shot generalization. To overcome these challenges, we propose the Structural-Semantic Aware Information Reduction (SAIR) framework. SAIR explicitly bridges this modality gap through three core components: (1) utilizing neuroscience-inspired strategies (edge extraction, depth estimation, and foveal blurring) to reduce visual redundancy and emphasize key structures; (2) employing parallel encoding architectures to systematically process multi-granular EEG features; and (3) applying a hierarchical contrastive objective for robust feature alignment. Extensive experiments on the Things-EEG dataset demonstrate that SAIR achieves state-of-the-art performance, with Top-1 accuracies of 66.9% and 17.5% in intra-subject and inter-subject settings, respectively, surpassing previous benchmarks by 8.8% and 3.8% absolute percentage points. Code is available at https://github.com/ThomasHC5/SAIR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1769_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ThomasHC5/SAIR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheHon_StructuralSemantic_MICCAI2026,
author = { Chen, Hongan AND Kong, Yan AND Shan, Caifeng AND Fang, Yuqi},
title = { { Structural-Semantic Aware Information Reduction for Asymmetric EEG-Visual Alignment } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The work is to focus on the EEG-Visual alignment. Given the rich information usually seen in images, one has to have relative compatible amount of information on both sides to bring the EEG-Visual alignment to reality. For instance, one has to suggest the focused area or the figural object(s) for the alignment.
- 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 method has clear advantage to other SOTA methods. The author(s) addressed the problem pretty well, essentially they compressed the visual information on the inputs (the novelty) compared to others who rely neural network to do it. They also introduce linear modulation for encoding EEG which let structure modulate the semantics.
- 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.
They only used one dataset, ATS used additional dataset (MEG) for generalization. They could have conducted analysis on temporal, spatial and spectral dynamics similar to NICE. Only intra-subject ablation, no inter-subject.
- 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
-
It is recommended to include a small paragraph or a table to summarize the notations used in this work.
-
For the encoder backbone, NICE already use Graph Attention (GAT) as EEG backbone so not new. Since they have three different embedding, compared to other works (NICE, ATS) they need three times more computation (assuming the encoder shared similar number of parameters) they should talk about the trade off.
-
For the training objective, its still pretty similar with most works which utilized infoNCE contrastive learning (they just extended it with multi embedding scenario). The experiments and discussion is brief, no training details too, low reproducibility from paper alone. They only used one dataset, ATS used additional dataset (MEG) for generalization. They could have conducted analysis on temporal, spatial and spectral dynamics similar to NICE. Only intra-subject ablation, no inter-subject.
-
- 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?
There are some novelty. However the experiments should be extended.
- 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
The paper presents a novel model to classify seen images from EEG recordings. The model uses neuroscience-inspired feature extraction techniques and a contrastive loss across multiple hierarchical stages. Results indicate that this model outperforms state-of-the-art in both subject dependent and subject independent evaluations.
- 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 key contribution lies in the novelty of employing image filters (edges, depth, foveal blurring) based on neuroscientific insights on the function of the visual system. The extracted image features are then aligned with EEG in a parallel, hierarchical pipeline using a contrastive loss, to mimic the different stages of visual processing in the brain. This architecture is conceptually sound and the justification and logic behind each component is well explained in the paper. The results are also quite good, with the model achieving superior performance compared to the state-of-the-art across almost all subjects.
- 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 main weakness is that there could be a few more results and evaluations of their model. Perhaps some results on the tuning of parameters, or on the impact of various types of information reduction strategies.
Another weakness is that the most up-to-date literature has not been shown. For example the Neurobridge paper (reference given below) has achieved 19% accuracy on the subject independent task, while this paper only achieves 17%. Actually, Neurobridge also uses some sort of neuroscience inspired preprocessing step for the images, similar to this paper. Zhang, Wenjiang, et al. “Neurobridge: Bio-inspired self-supervised eeg-to-image decoding via cognitive priors and bidirectional semantic alignment.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40.No. 21.2026.
- 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 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
- It would be better to remove “the source code will be available” from the abstract. It should be added somewhere else (perhaps under Implementation Details)
- The first contribution states “We identify…”. The paper does not report any structured tests done to identify and validate their claim of the limitations of current models. This should not be listed as a contribution.
- The contributions listed in the introduction are different to those in the abstract. It is better to have them be the same, perhaps with more detail in the introduction
- The idea of using depth maps and other features and applying a contrastive loss to align EEG with image features seems to be based on prior work. For example, CogCap (reference 21 in this paper) has used a similar method for image generation, although the results are much worse than this paper. The authors should mention this paper when explaining their methods.
- I suppose foveal blurring only would work when the subject is centred. How would this model work with images with subjects off-centre? Will the model perform worse than state-of-the-art in such cases?
- Error in Figure 2 (c) the 2 EEG Encoders should be changed to 2 CLIP encoders
- Topology agnostic cross-subject mode seems to be discarding what was learnt by GAT. How do the authors ensure this does not cause instability or other issues? Also, what about subject-specific patterns in time, not electrode space?
- Why don’t the authors use independent lambda weights for each term in eq 3?
- Why are there 4 repetitions/image in the train subset but 80 repetitions / image in the test subset?
- It is difficult to read Fig 3 (b) ablation study. My opinion is that barplots are usually not a great method to display such data visually. For these straightforward results (ablation accuracies), it probably better to present as a table.
- Table 1 top-1 accuracies for ATS are different to their published results (e.g., average 58.1 reported here, 60.9 reported in ATS paper). Why is this?
- 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?
I believe the authors provide a strong neuroscience based foundation for their model, and clearly explain the reasons for their architectural choices. The model also achieves significant improvements over the state-of-the-art, although some very recent papers may have slightly better results.
- 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
The authors propose a framework namely, Structural-Semantic Aware Information Reduction (SAIR) to address semantic mismatch where image backgrounds overshadow the primary subject by using neuroscience inspired strategies including edge extraction, depth estimation and foveal blurring. Authors later apply hierarchical contrastive objective for feature alignment and show significant improvement in performance on the Things-EEG dataset, supported by both quantitative results and qualitative analysis.
- 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 paper is well-written with clear motivation and, comparison with existing methods. the authors provide a comprehensive set of ablations and qualitative results to validate their claims.
A particularly interesting example is shown in Figure 3 (row 4), where for the query “crab”, SAIR does not retrieve the exact object but still returns semantically related items such as mosquito and spider. While imperfect, this behavior highlights that the model captures meaningful feature-level similarities, indicating robust retrieval capabilities beyond exact matching.
It would be interesting to explore its extension to open-vocabulary retrieval settings, which seems feasible given the design.
- 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 use of foveal blurring is compelling, but it raises a potential limitation: since foveal emphasis is typically centered, the method may struggle when the primary object lies in the periphery of the image. Clarification or empirical analysis of such cases would strengthen the work.
2.Additionally, there is some ambiguity in the architectural description. Figure 2(c) suggests that depth and foveal processing are applied via the EEG encoder, while the text (Page 5, Paragraph 2) implies that CLIP processes all reduced images (edge, depth, fovea). It is unclear whether CLIP is applied only to edge features or to all modalities.
a. If this is a typo, the authors should correct it for clarity. b. If intentional, a justification for this design choice would be valuable.
- 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 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
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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a well-motivated and thoughtfully designed approach to a known limitation in EEG-visual alignment due to background dominance . The proposed SAIR framework is conceptually clean, biologically inspired, and empirically effective. The combination of multiple structural cues with hierarchical contrastive learning is both novel and practical. Despite minor clarity issues and a potential limitation regarding spatial bias in foveal processing, the overall contribution is strong. The experimental validation is convincing, and the qualitative results provide meaningful insight into model behavior. Given this case, I recommend a strong accept.
- 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
Author Feedback
We sincerely thank the AC and Reviewers for constructive feedback and for recognizing the novelty of our SAIR framework. We are thrilled by the positive evaluation and address main concerns below. (Q-question A-additional comment) 1.ATS Accuracy Mismatch (R2A11) We clarify that the ATS paper’s main table reports results from a single seed (seed=0), as evidenced by their accuracies ending in 0.5% (200-way) and confirmed in their supplementary material. For rigorous evaluation, we report the average performance across 5 runs (seeds 0-4). Our reproduced 5-seed average for ATS (58.1%) closely aligns with the only overall 5-seed average provided in their supplementary material (59.07%). 2.Comparison with NeuroBridge (R2Q2) NeuroBridge is an excellent concurrent work. Since it was officially published online in Proceedings of AAAI on 16 March 2026, it was not included as a baseline. We acknowledge their inter-subject advantage (19% vs. 17%). However, our paradigms differ fundamentally: NeuroBridge simulates cognitive-level perceptual variability via statistical data augmentation (e.g., noise, mosaic), while SAIR extracts structural features (Edge, Depth, Foveal) related to the early visual system for structurally constrained alignment. We view these two contemporaneous works as highly complementary, and exploring the fusion of SAIR with NeuroBridge would be an exciting future direction. 3.Component Novelty (R1A2&A3) We deliberately utilized standard GAT and InfoNCE for stable feature extraction and fair benchmarking. Our core novelty lies not in inventing components, but in the system-level SAIR framework. By uniquely integrating biological structural cues into a multi-embedding contrastive paradigm, we provide a novel solution to the visual-EEG semantic mismatch problem. 4.Topology-agnostic Mode Stability (R2A7) We honestly acknowledge that discarding GAT in cross-subject settings is an engineering compromise to prevent overfitting to subject-specific spatial artifacts, as stated in Sec. 2.3.In our work, model stability relies on the robust shared representation space established by CLIP, which warrants further investigation and improvement in future work. 5.Foveal Blurring on Off-center Objects (R2A5&R3Q1) Things-EEG was collected using the standard Rapid Serial Visual Presentation paradigm, where a central fixation cross continuously guides the subjects’ attention to the screen center. Thus, under this protocol, central blurring is an accurate biological reflection rather than a limitation. We will explicitly detail this to resolve ambiguity. Extending our work to free-viewing settings is a direction for future work. 6.Computation Trade-off (R1A2) We acknowledge that the three embeddings triple the resource requirements. However, this is a worthwhile trade-off: explicitly extracting biologically plausible cues yields significant performance gains. Moreover, since the three branches compute in parallel, actual inference latency remains low. 7.Independent λ Weights (R2A8) Independent tuning of λ weights in Eq. 3 yielded only marginal gains over equal weighting. We chose equal weights to prioritize simplicity. 8.Training Details & Reproducibility (R1A3) Our core training configurations are already detailed in Section 3.We apologize if they were easily overlooked. To improve readability, we will include additional minor hyperparameters. Code and weights will be released soon. 9.Ambiguity in Fig. 2(c) (R2A6&R3Q2) We sincerely thank the reviewers for catching this. “EEG Encoder” was a typo and has been corrected to “CLIP Encoder”. 10.Additional Experiments & Presentation (R1&R2&R3) Due to strict page limits, including extensive analyses (MEG dataset, spatial-temporal dynamics and extra ablations) are infeasible. We will explore these valuable suggestions in our future work. Regarding other presentation improvements and minor corrections, we will do our utmost to maximize the clarity of the camera-ready version.
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.
The paper is well motivated and the experiments are convincing. It is recommended to compare and discuss with more recent literature in this area.