Abstract

Animal models, particularly rats, play a critical role in seizure research for studying epileptogenesis and treatment response. However, progress is limited by the absence of datasets with precise temporal annotations and standardized evaluation protocols. Existing animal behavior datasets often provide limited accessibility, coarse labeling, and insufficient temporal localization of clinically meaningful events. To address these challenges, we introduce RatSeizure, the first publicly available benchmark designed for fine-grained seizure behavior analysis. The dataset is curated from long-form recordings and segmented into 10-second clips with dense annotations, including seizure-related action units and accurate temporal boundaries, enabling both behavior classification and temporal localization tasks. In addition, we propose RaSeformer, a saliency-context Transformer for temporal action localization that emphasizes behavior-relevant temporal context while suppressing redundant cues. Experimental results on RatSeizure demonstrate that RaSeformer achieves strong performance and provides a competitive reference model for this challenging setting. We further establish standardized dataset splits and evaluation protocols to promote reproducible benchmarking. Dataset is available at https://github.com/UA-CVML/RatSeizure.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0427_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/UA-CVML/RatSeizure

Link to the Dataset(s)

https://github.com/UA-CVML/RatSeizure

BibTex

@InProceedings{TsaTin_RatSeizure_MICCAI2026,
        author = { Tsai, Ting Yu AND Yu, An AND Lee, Lucy AND Ye, Felix X.-F. AND Shin, Damian S. AND Kao, Tzu-Jen AND Li, Xin AND Chang, Ming-Ching},
        title = { { RatSeizure: A Benchmark and Saliency-Context Transformer for Rat Seizure Localization } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces RatSeizure, a benchmark for fine-grained seizure behavior analysis in rats, with clip-level annotations for behavior classification and temporal localization. The authors also propose RaSeformer, a Transformer-based temporal action localization model with saliency-guided sparse attention.

  • 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 is generally well-written and clearly organized. 2.The application is valuable. Fine-grained seizure behavior recognition in Alzheimer’s disease models can potentially address unmet clinical needs in preclinical behavioral assessment.

  • 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 benchmark contribution is potentially useful, but the current dataset design and presentation still raise concerns regarding scale, annotation quality, and benchmarking rigor.

    • The dataset contains only 794 clips from 9 long recordings, while the effective number of distinct animals is not clearly specified. This limits confidence in subject diversity and the generalizability of the benchmark.
    • The annotation granularity could be better justified. Since the benchmark is positioned as fine-grained seizure behavior analysis, the paper should more clearly explain the value and limitations of the current annotation setting. Stronger temporal supervision or a clearer justification of the chosen granularity would strengthen the benchmark contribution.
    • The positioning relative to existing rodent-related resources is not fully clear. In particular, the paper would benefit from a clearer discussion of how RatSeizure differs from broader rat behavior datasets such as Rat7M, as well as prior seizure-related rodent datasets such as RodEpi. Accordingly, the claim of being the first publicly available fine-grained seizure benchmark should be more carefully contextualized.
    • The data split strategy requires clearer explanation. Since the benchmark is constructed from a small number of long recordings, it is important to clarify whether clips from the same recording can appear in both training and test sets, and how potential temporal or recording-level leakage is avoided.

    2.The methodological novelty appears somewhat limited. RaSeformer mainly combines existing components, including I3D features, a Transformer encoder, sparse attention, and a multi-scale temporal pyramid. The proposed per-head Top-K pruning seems to be a relatively modest extension rather than a substantial methodological advance.

    3.The experimental validation is not sufficiently comprehensive.

    • The empirical comparison is limited to four TAL baselines, all from 2022 to 2024.While these are relevant methods, the evaluation would be stronger with broader coverage of more recent developments, including cited methods such as AdaTAD [13] and SNAG [16], as well as a brief discussion of newer work from 2025 and 2026.
    • It also lacks comparison with skeleton-based approaches such as VSViG[27], which are relevant in biomedical video analysis due to their efficiency and robustness.
    • The evaluation reports only mAP, without additional metrics such as recall, F1 at different IoU thresholds, or per-class analysis. Given the likely class imbalance in seizure behaviors, a more detailed evaluation would be valuable.
    • The paper also lacks computational cost analysis, such as runtime, complexity, or FLOPs.
  • 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.

    (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?

    This paper introduces a useful benchmark and a corresponding method for seizure behavior recognition in rat models, addressing a practically relevant problem. The manuscript is generally clear and well organized. However, several limitations discussed in Q7 prevent a higher score. Overall, I consider the paper slightly above the acceptance threshold and therefore recommend Weak Accept.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal reasonably addresses my main concern about possible data leakage by clarifying that the split was done before augmentation and at the source-context level. I still have reservations about dataset scale, annotation reliability, and limited evaluation, but these are not fatal. I would maintain my Weak Accept recommendation.



Review #2

  • Please describe the contribution of the paper

    Creating a public seizure benchmark is the main contribution of the paper. The action-unit annotations and temporal boundaries make the dataset more useful than prior rodent behavior datasets, which are often too coarse to support meaningful localization. RaSeformer is a reasonable accompanying model, but I see the benchmark as the more important part of the paper.

  • 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 strongest part of the paper is the dataset itself. A public benchmark with seizure-specific action units, temporal boundaries, and a clear evaluation protocol would be genuinely useful. The paper also does a good job explaining how RatSeizure differs from earlier animal behavior datasets, and the action-unit table helps make the label space easy to understand.

    2.The clinical motivation is convincing. Automatically localizing seizure behaviors in preclinical rodent videos is clearly relevant for epilepsy research, and this feels like an important problem that has not received much attention in medical video analysis.

    3.The modeling approach is also rigorous. The preprocessing choices, especially mirror masking and ROI cropping, seem well matched to the recording setup. The sparse local temporal attention design also makes sense for short, repetitive behaviors where fine-grained temporal cues matter more than long-range context. The pipeline figure was easy to follow.

    4.The experiments show consistent improvements over the baselines the authors consider. The ablation study also suggests that both query-dependent saliency scoring and head-specific Top-K selection are helpful. The gains are not huge, but they point in the same direction across the reported subsets.

  • 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 with the evaluation setup. The dataset is built from nine long recordings, the clips are extracted with a 10-second window and 5-second stride, and augmentation is used to deal with class imbalance. That immediately raises a question: how exactly was the split done? If the split is only at the clip level, then overlapping clips from the same recording could easily end up on both sides, and augmented variants of the same source clip could do the same. The paper does not make this clear enough, and for a benchmark paper that is a critical problem

    I also do not think the annotation-quality evidence is strong enough yet. The authors describe a multi-expert review process, which is good, but they also say that intermediate annotation states were not kept, so there is no inter-rater agreement analysis. For a dataset whose value depends heavily on temporal boundaries, that is a real weakness.

    On the method side, the novelty is moderate. Local-window sparse attention with Top-K pruning is a sensible design, but it does not strike as a major step beyond the general family of sparse/local temporal attention models. The model is good; however I do not think it is the part of the paper doing the heavy lifting.

    The comparison framing is overstated. The paper claims comparison against “state-of-the-art temporal action localization methods,” but this is not quite right. TriDet and ActionFormer are strong, standard baselines. TadTR is older and no longer a convincing SOTA reference. TE-TAD is a strong query-based baseline, but even its own paper frames it as outperforming previous query-based detectors while being competitive with SOTA, which is a narrower claim. More recent work such as AdaTAD and DiGIT, as well as the OpenTAD benchmarking study, suggests that the field has already moved beyond the exact baseline set used here. So I think “strong representative baselines” would be fair, but “state-of-the-art” is overstated

    The experimental validation is narrower than I would expect for MICCAI. The paper reports aggregate mAP, but not a fuller set of metrics such as mAP across individual IoU thresholds, F1 score, Edit score, or accuracy, all of which would help clarify whether the gains reflect better localization, better boundary quality, or simply better overall detection. In addition, there is no per-class analysis, no onset/offset error analysis, no confidence intervals or repeated-split evaluation, and no cross-recording generalization study. The manuscript also cites stronger recent TAL works such as AdaTAD and SNAG but does not compare against them. Given the small and specialized dataset, the empirical claims feel less settled than the paper suggests.

  • 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.

    (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?

    My overall score is driven by a mix of genuine interest in the benchmark contribution and fairly serious concerns about the current validation. The strongest part of the paper is the dataset: a public rat seizure benchmark with action-unit labels and temporal boundaries could be useful, and this is an under-served problem in medical video analysis. The paper is also readable, the pipeline is easy to follow, and the proposed model is a reasonable reference baseline for the task. The main reasons I scored it below the acceptance bar are on the experimental side. First, the split protocol is not described clearly enough for a dataset built from only nine long recordings with overlapping 10-second windows and augmentation. I was left uncertain whether overlapping clips or augmented variants from the same source material could appear across train and test, and that matters a lot for how much confidence one should place in the reported gains. Second, the annotation-quality evidence is limited. The paper describes a multi-expert review process, but does not provide inter-rater agreement or similarly strong evidence for temporal boundary reliability. Third, the model contribution feels more incremental than the paper suggests. RaSeformer seems like a sensible combination of known ingredients rather than a major methodological advance. I was also not fully convinced by the empirical framing. The paper describes the comparison set as “state-of-the-art,” but that overstates the case. TriDet and ActionFormer are strong representative baselines, but TadTR is older, and the omission of more recent high-end TAD methods makes the comparison feel less current than the wording implies. More broadly, the evaluation is narrower than I would expect for a temporal segmentation paper. The paper reports aggregate mAP, but not a fuller metric breakdown such as mAP at different IoU thresholds, F1 score, Edit score, or accuracy, and there is no per-class analysis, no cross-recording generalization study, and no repeated-split evaluation. Given the size and specificity of the dataset, those missing pieces make the empirical claims feel less settled than the paper suggests. Overall, I see this as a promising benchmark paper with a credible application and a useful potential resource, but not yet one that is fully convincing in its current form.

  • 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 claim of “a novel AI-driven framework for the automated assessment of microanastomosis performance” is an overstatement. The paper is positioned more as a methodological contribution than what it actually delivers, which is closer to a dataset contribution with a relatively standard modeling pipeline applied on top. The omission of more recent high-end TAD methods from the comparison makes the experimental evaluation feel less current than the framing implies. More broadly, the evaluation is narrower than I would expect for a temporal segmentation paper, both in terms of baselines and in the range of metrics and ablations reported. For these reasons, I believe the paper requires major revisions and I recommend rejection in its current form.



Review #3

  • Please describe the contribution of the paper

    This paper introduces RatSeizure, a new publicly available benchmark dataset for fine-grained analysis of rat seizure behaviour, with dense temporal annotations and standardised evaluation protocols. The dataset consists of curated clips from long-form recordings with precise temporal boundaries and clinically motivated action units, enabling both classification and temporal action localisation tasks.

    In addition, the paper proposes RaSeformer, a saliency-context Transformer for temporal action localisation. The model incorporates a sparse attention mechanism using per-head Top-K pruning within local temporal windows to emphasise behaviour-relevant context while suppressing redundant information.

    The authors also establish standardised dataset splits, evaluation protocols, and benchmark results against existing temporal action localisation methods, demonstrating improved performance on the proposed dataset.

  • 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.

    RatSeizure addresses a clear gap by providing the first publicly available dataset with dense temporal annotations for seizure behaviour. Comprehensive dataset design, combined with a multi-stage expert annotation and verification pipeline, enhances label quality. Includes handling of class imbalance and augmentation strategies. Regarding the method, a clear three-stage design (preprocessing, feature extraction, temporal modelling). Effective use of sparse attention for efficiency and relevance.

  • 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 scale is relatively small, with only 794 clips derived from 9 recordings, covering only the light phase (7 am-7 pm) and male adults. This may limit generalisation and robustness. The dataset has limited diversity with a single experimental setup and controlled environment.

    The model has moderate novelty as RaSeformer builds on existing Transformer-based TAL frameworks with incremental improvements (Top-K pruning, local attention). The main novelty lies in the dataset contribution rather than the model design. Evaluation is limited to in-dataset performance with no cross-dataset generalisation or transfer evaluation. Heavy augmentation of classes to balance them may influence performance comparisons; analysis of real versus augmented performance could be expanded.

  • 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

    To further strengthen the work, the authors are encouraged to:

    • Expand discussion on dataset limitations and generalisation potential.
    • Provide additional analysis on model efficiency (runtime, FLOPs).
    • Include cross-dataset or transfer experiments if possible.
    • Provide a deeper analysis of the effects of augmentation.
    • Explore scaling the dataset in future work to enhance impact.
  • 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?

    This paper makes a strong contribution by introducing a novel, benchmark dataset for fine-grained analysis of seizure behaviour with dense temporal annotations. The dataset fills an important gap and has clear clinical and research relevance.

    While the proposed model is moderately novel, it is well-designed and performs well compared to existing methods. The combination of dataset contribution, solid experimental validation, and clear presentation makes this work valuable to the MICCAI community.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors also appropriately recalibrate several overstated claims and clarify the intended contribution of the benchmark. While methodological novelty remains moderate and evaluation breadth is still somewhat limited, I view the work as a valuable benchmark/resource contribution with a reasonable accompanying baseline.



Author Feedback

We thank the reviewers for the thoughtful feedback. Below, we summarize and address each point.

P1-Split Integrity The dataset was split before augmentation: source clips were assigned to train or test, and augmented variants inherited the same split. Thus, no original clip or augmented variants appear across partitions, preventing augmentation-induced leakage. RatSeizure adjacent clips may share temporal content. To prevent overlap-related leakage, we split at the source-context level rather than by randomly mixing clips. Clips from the same video or seizure episode were kept in the same partition, so overlapping, adjacent, or source-related clips do not cross the train/test boundary. We report dataset distributions in Fig. 1 and separately provide full, non-augmented, and augmented test results to ensure performance on original data is not obscured.

P2-Annotation Reliability RatSeizure used sequential expert adjudication rather than parallel independent annotation. One expert marked temporal AU segments, and three experts refined labels and boundaries to reach consensus. Since intermediate snapshots were not retained, formal IAA cannot be computed retrospectively; we will clarify this limitation. As QC evidence, Expert 1 revised about 20% of clips, while Expert 2 added only 5% corrections, suggesting most discrepancies were resolved early with limited later refinement. To assess temporal-boundary reliability, we will add a small blinded re-annotation audit. A subset of clips will be independently re-labeled by two experts, with boundary consistency reported using temporal deviation and/or temporal IoU. We will distinguish these audit statistics from the formal IAA. We will also clarify annotation granularity: the 5-second criterion defines seizure episodes for inclusion and clip extraction, while shorter AUs within valid episodes and concurrent AUs are annotated as multi-label temporal segments and evaluated independently.

P3-Dataset Scope and Generalization Dataset details are provided in Secs. 3 and 6.While these factors may affect generalizability, they support a controlled and reproducible benchmark for expert-annotated seizure AU localization. Nine videos correspond to nine distinct rats, matching the localizing task of which seizure-related AU occurs and when, rather than modeling multi-animal interaction or identity tracking. We will refine positioning: RatSeizure is not the first public rodent dataset overall, but it is the first public benchmark for fine-grained temporal localization of rat seizure behaviors with AU labels and boundaries. Rat7M lacks seizure-specific AU labels and boundaries, RodEpil focuses on binary seizure detection, and VSViG requires pose estimation with a different input setting.

P4-Method Positioning The method, evaluation, and ablation are described in Secs. 4 and 5.We agree RatSeizure is the primary contribution and will frame RaSeformer as a task-adapted reference model rather than the main novelty. Results indicate that RaSeformer performs best among compared methods across full, non-augmented, and augmented test sets. The ablation supports query-dependent saliency and per-head specialization. We also agree that “SOTA” overstates the baselines and will revise it to “strong representative TAL baselines.” Our goal is to establish a reproducible benchmark, not to exhaustively cover the TAL literature. Newer methods like AdaTAD/SNAG will be discussed as future extensions without overclaiming.

P5-Expanded Evaluation We agree that finer metric breakdowns and efficiency reporting would improve interpretability. Since rebuttal policy precludes new results, we will clarify them as future extensions. We follow standard TAL evaluation, using avg. mAP over tIoU 0.3-0.7 as the primary metric; per-threshold scores were computed but not reported separately. Full, non-augmented, and augmented results are reported separately, supporting our reproducible benchmark with a competitive reference model.




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 receives two week accept and one week reject. The concerns about the benchmark details are serious but fixable.

    Points to address:

    • data split protocol
    • annotation quality
    • evaluation strengthening
  • 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.

    Both reviewer 1&3 find the rebuttal address their original concerns on dataset leakage. Despite the remaining concerns about the dataset size and annotation quality, it is a good contribution to the community as a benchmark paper.



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.

    The paper introduces RatSeizure, a publicly released benchmark for fine-grained temporal localization of rat seizure behaviors with multi-expert annotations, and a saliency-context transformer as a reference model. The main concerns were split integrity under heavy augmentation, annotation reliability without formal inter-annotator agreement, dataset scope, and the model’s moderate methodological novelty. The rebuttal clarifies that the split is performed at the source-context level before augmentation, and that augmented variants inherit the same partition, thereby eliminating augmentation- and overlap-induced leakage. The authors also acknowledge the absence of formal IAA, explain the sequential adjudication protocol, commit to a blinded re-annotation audit, recalibrate the benchmark’s positioning, and reframe the model as a task-adapted reference rather than the primary contribution. Two reviewers explicitly endorsed acceptance after the rebuttal; R2 maintained Reject due to moderate methodological novelty and limited breadth of evaluation, but the dataset and benchmark contribution stand on their own, and the methodological component is appropriately scoped. The camera-ready should incorporate the revisions.



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 paper is recognized by all reviewers as making useful contributions mainly to the dataset, together with a reasonable modeling pipeline. Meanwhile, there has been concerns on the lack of major technical contributions, and the issues related to over-claims, among others. The authors must follow the reviews very carefully, and address the major issues/concerns in the camera ready, if it is accepted by MICCAI.



back to top