Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Intraoperative Adverse Events (IAEs) detection is critical for improving surgical safety, with bleeding being among the most frequent events across many surgery types. Existing methods struggle to distinguish bleeding IAE from visually similar residual blood due to limited temporal reasoning. Moreover, modeling long surgical videos while preserving fine-grained temporal dynamics remains computationally challenging. We propose DBT-Bleed, a dual-branch multi-scale temporal modeling framework disentangling bleeding and normal representations using layer-wise temporal adapters for short- and long-term bleeding progression. To efficiently process long surgical videos without sacrificing fine-grained temporal information, we introduce HiRED, a Hierarchical Entropy-Driven frame selection strategy that retains temporally informative segments while removing redundancy. Experiments on the MultiBypass dataset demonstrate gains of 6.53% in F1, 5.62% in Recall and 9% in MCC values for bleeding IAE detection, consistently outperforming video-level baselines. Additionally, we evaluate cross-procedure generalization on a newly curated dataset from a different surgical procedure type, where DBT-Bleed demonstrates robust transferability by achieving gain of 6% in F1 and 8% in MCC under zero-shot setting. To support this evaluation, we introduce EndoPit-IAE, an Endonasal Pituitary Surgery dataset annotated for IAEs, representing the first IAE-annotated dataset in neurosurgery. Code will be made publicly available upon acceptance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4663_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/jinlab-imvr/DBT-Bleed

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MisSud_DBTBleed_MICCAI2026,
        author = { Mishra, Sudhanshu AND Xu, Jialang AND Ang, Jensen AND Mazomenos, Evangelos B. AND Ang, Beng Ti AND Jin, Yueming},
        title = { { DBT-Bleed: Dual-Branch Temporal Modeling with Key-Frame Selection for Surgical Bleeding Detection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16892},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors describe a clip-based model for detecting intraoperative adverse events. In particular, the claim is to differentiate between adverse events and residual blood in the scene, which have similar visual signatures. To approach this, the work proposed a frame-selection strategy which is entropy-based. Following this, a temporal adapter puts together embeddings from multiple temporal resolutions. Evaluation is additionally performed on a curated neurosurgical 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.
    • The problem that the paper deals with is of high clinical relevance, challenging, and a novel application area.
    • The work introduces a curated surgical dataset containing adverse events which is useful for the community.
    • Encoding different temporal resolution is a good approach that can combine different kinds of context which contributes to the procedural understanding that surgical video requires.
  • 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.
    • One of the major claims of the paper is the aspect of differentiating non-residual bleeding and adverse events. This is from a clinical perspective, a really hard problem. The paper does not analyse this in sufficient detail. For example, how are these events formally defined and classified, and how many of them were correctly hit or missed by the model? How does the work differentiate between annotation errors and residual bleeding? Given that the authors place focus on this aspect, the analysis and methodology does not sufficiently back up this claim.
    • The work does not report precision, which could be misleading because high recall and F1 can indicate that the model could be biased by the negative examples, considering adverse events are rare and heavily imbalanced. (This is also supported by the low MCC values).
    • The paper classifies “bleeding in any frame” as a positive bleeding label for the clip. A justification is needed for this design choice. A more natural choice would be to use a majority vote. This design choice is more prone to inflated numbers.
    • Frame selecting based on red channel information makes sense but does not address the problem of adverse vs residual bleeding.
    • How many annotators were used for annotating the in-house dataset? How was the protocol defined and validated? What was the inter-rater agreement? These details are needed to support the description of the data.
    • There are quite a few surgical VLP baselines that are missing from the comparison.
  • 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 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
    • Consider partitioning the test set into residual and active bleeding, which could potentially enable a quantitative analysis of what the paper claims.
    • Present in more detail the annotation protocol and inter-rater agreement (Cohen’s kappa) and the relationship between VIBe and the SEVERE scale as used in the MBP140 dataset.
    • In case of imbalance dataset, precision is important, and similarly the AUPRC metric as well. Consider adding these.
    • Pick majority vote
  • 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?
    • As detailed in the weakness section, the central claim of the paper is not sufficiently substantiated. The proposed components do deliver an improvement, but the mechanism is not sufficiently looked into , in order to address the main claim of the paper.
    • Important surgical VL baselines are missing, and pertinently reporting of precision and AUPRC metrics. The difference with MCC indicates modest performance under class imbalance.
    • Evaluation dataset is not presented in detail. Important details such as annotator diversity, agreement, and protocol is missing, which reduces the ability to evaluate model performance. Due to the reasons detailed above, the paper in its current version is not yet ready for publication in the MICCAI conference.
  • 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 have presented arguments that address my concerns. It is important that the claims about residual bleeding are tempered as qualitative and the difference explained. Although there are some limitations, I believe the work is a valuable contribution to a relevant problem. Therefore, I change my vote to accept.



Review #2

  • Please describe the contribution of the paper

    1) DBT-Bleed: a dual-branch multi-scale temporal modeling framework disentangling bleeding and normal representations using layer-wise temporal adapters for short- and long-term bleeding progression. 2) HiRED: a Hierarchical Entropy-Driven frame selection strategy that retains temporally informative segments while removing redundancy, to efficiently process long surgical videos without sacrificing fine-grained temporal information.

  • 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 robust temporal modeling to distinguish bleeding IAE from visually similar residual blood. 2) The modeling of long surgical video sequences without losing fine-grained temporal dynamics. 3) The proposed methods were evaluated on a publish dataset MultiBypass via comparison and ablation studies. A new dataset is also established.

  • 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 application scope is not clear. Although the paper stated “it is crucial to timely detect bleeding to enable prompt intervention and improve surgical outcomes”, the real-time perforamce is not revealed. The framework requires a long video sequence to distinguish bleeding more reliably, which also leads to a larger latency of IAE detection during surgery. 2) The proposed methods is only compared to related CLIP/Mamba based methods, but not compared to the two methods evaluated on the same public dataset MultiBypass. 3) Fig 2(b) is not clear. The modules S and A are not introduced. The dashed arrows are not informative. 4) The higher scores on EndoPit-IAE than those on MultiBypass. The explaination “is attributed to higher resolution of EndoPit-IAE videos with high color contrast” is conjectured and not proved.

  • 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 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?

    Pros: The proposed method is capable to distinguish bleeding against residual blood, via key frame selection and temporal modeling.

    Cons: 1) The application scope is not clear. Although the paper stated “it is crucial to timely detect bleeding to enable prompt intervention and improve surgical outcomes”, the real-time perforamce is not revealed. The framework requires a long video sequence to distinguish bleeding more reliably, which also leads to a larger latency of IAE detection during surgery. 2) The proposed methods is only compared to related CLIP/Mamba based methods, but not compared to the two methods evaluated on the same public dataset MultiBypass.

  • 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.

    Although this work is well motivated by the demand for bleeding IAE detection, the technical route, which ustil is still not convincing in terms of practical applications.



Review #3

  • Please describe the contribution of the paper

    The paper presents DBT-Bleed, a dual-branch multi-scale temporal modeling framework for the detection of intraoperative adverse events (IAEs), specifically focusing on bleeding. The novel contribution lies in the ability to effectively distinguish between bleeding events and residual blood, which can appear visually similar in surgical videos. By using layer-wise Multi-scale Temporal Adapters (MTA), the method captures both short- and long-term progression of bleeding, improving the accuracy of detection. The framework also introduces HiRED, a Hierarchical Entropy-Driven frame selection strategy that allows for efficient processing of long surgical videos while maintaining fine-grained temporal information. The approach is validated using two datasets, MultiBypass and EndoPit-IAE, demonstrating strong performance in both general bleeding detection and cross-procedure generalization.

  • 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 proposes a dual-branch multi-scale temporal framework (DBT-Bleed) capable of performing both short-term and long-term temporal reasoning, enabling the model to distinguish subtle bleeding intraoperative adverse events from visually similar residual blood. (2) The paper introduces a hierarchical entropy-driven frame selection strategy (HiRED), which efficiently processes long surgical videos by selecting video frames that carry key temporal information, significantly improving computational efficiency. (3) The method demonstrates strong zero-shot generalization on the newly constructed EndoPit-IAE dataset, indicating that the model can handle different types of surgical scenarios without additional training.

  • 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 focuses solely on binary bleeding detection and does not explore the detection of multiple types of IAEs, which may limit its clinical applicability. Extending the framework to support multi-label IAE detection would broaden its utility. (2) Although the paper demonstrates cross-procedure generalization on the EndoPit-IAE dataset, testing on a wider variety of surgical procedures would provide stronger evidence of the model’s versatility and robustness. (3) The dual-branch framework and key-frame selection strategy improve performance but may introduce complexity and computational overhead. A more detailed discussion on the model’s computational requirements and potential trade-offs would help evaluate its practicality for real-time applications. (4) The paper’s results are primarily based on videos of a specific length (300 frames), and the method’s performance may vary with videos of different lengths. Further exploration of how video length affects detection and performance would be valuable for clinical deployment.

  • 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 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.

    (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?

    (1) The paper introduces DBT-Bleed, a dual-branch multi-scale temporal framework for surgical bleeding detection, which effectively addresses the challenge of distinguishing between actual bleeding and residual blood during surgeries. This innovation, combined with HiRED, a novel hierarchical entropy-driven frame selection method, provides a significant advancement in the field of intraoperative adverse event (IAE) detection, particularly with its ability to handle long surgical videos while retaining fine-grained temporal dynamics. (2) The paper clearly defines the problem of detecting bleeding IAEs and introduces a method that can distinguish between actual bleeding and non-bleeding residual blood, which is a significant challenge in surgical video analysis. (3) The framework holds great potential for clinical applications, particularly in enhancing surgical safety by providing real-time detection of bleeding incidents. This can significantly reduce complications and improve outcomes for patients. Moreover, the introduction of EndoPit-IAE, a neurosurgical dataset annotated for IAEs, opens up new avenues for further research in this domain. (4) While the paper presents a novel combination of existing techniques (like CLIP-based models and multi-scale temporal adapters), the underlying methodologies themselves are not groundbreaking, as similar approaches have been explored in other video anomaly detection tasks. More emphasis could be placed on explaining how these methods are particularly tailored to the unique challenges of surgical bleeding detection.

  • 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.

    N/A



Author Feedback

We are grateful to R1,R2,R3,MR for constructive feedback. Point-by-point answers follow. Q1-Residual vs Active bleeding-R1&MR We thank R1 for recognising this clinically challenging task. For analysis: We follow the MultiBypass IAE protocol[1] for bleeding-IAE(active bleeding) definition. Bleeding-IAE is unintended bleeding at the surgical site from mechanical/thermal incision; residual blood (minor, settled, pooled, preparation-related, or post-event unresolved) is labeled normal. We will clarify the definitions in Intro & Fig. 1.Because MultiBypass neither annotates residual blood separately nor flags annotation errors, residual-vs-IAE quantification would require test-set relabeling, which is beyond scope. Under this constraint, we present the first dedicated bleeding-IAE study on MultiBypass following its IAE-label release. Table 2 demonstrates quantitatively: higher MCC reflects lower False-Positive(FP) rate, and improvements across all metrics show our method is best at identifying bleeding-IAEs. Visual analysis (Figs. 3-4) shows baselines misclassify residual blood as IAE, yielding high FP rates (Fig. 3). For method: Bleeding-IAE is inherently temporal, and our framework exploits this. HiRED is addressed in Q2.MTA captures multi-scale temporal cues: low-level color/intensity shifts and deeper instrument–tissue patterns reflecting consistent blood flow across frames, helping distinguish bleeding-IAEs from residual bleeding. Q2-Role of HiRED-R1&MR HiRED is a frame selector, not the final discriminator. It reduces redundant residual-blood frames and promotes high red-channel entropy frames, more common in bleeding-IAE despite distribution overlap. We will add entropy-distribution plots for IAE, residual, and normal frames. Ablation study (Table 3) confirms performance improvement due to HiRED. Q3-Evaluation metrics-R1&MR F1/Recall are intentionally reported for the target bleeding-IAE class. They use TP/FP/FN (not TN), so are not inflated by negative clips. Precision is derivable from the F1–recall formula; DBT-Bleed achieves the highest precision at 55.26, vs. second-best 50.00.MCC accounts for all confusion-matrix entries under imbalance; while absolute MCC reflects task difficulty, DBT-Bleed achieves the best MCC, showing stronger discrimination of IAE from residual/normal cases. DBT-Bleed also improves AUPRC over MadCLIP (61.6 vs. 56.29), which we will add. Q4-Real-time analysis-R2&R3 On a single RTX A5000, DBT-Bleed runs at 204.8 fps. It uses a 300-frame (~12s at 25fps) context before first prediction, refreshing every ~8s; latency is bounded by the temporal window - a deliberate accuracy/latency trade-off supported by our N-sweep ablation. Q5-Annotation protocol on inhouse-dataset-R1&MR Annotated by 3 raters: one consultant surgeon (>10 yrs) and two consultant-trained trainees. Trainee labels were reviewed by the consultant; discrepancies were resolved via joint re-review and consensus. Average inter-rater discrepancy was <0.5% per video. Q6-Clarifications on majority voting, etc-R1&R2&R3&MR R1&MR: Since IAEs are sparse, majority voting with N=300 requires ≥151 positive frames, missing short IAEs (383/474 training bleeding events are <150 frames). Metrics aren’t inflated (see Q3). R2: We tested LEMON-FM as SEDMamba’s visual encoder without improvement. As surgical VLPs are emerging and yielding no gains, broader SurgVLP integration is left for future work. For two MultiBypass-specific methods, code/predictions were unavailable, so reproduction was not possible. R2: We will revise the EndoPit-IAE discussion cautiously: higher resolution may partly explain higher absolute scores vs. MultiBypass(unverified). Our supported claim is comparative: under the same zero-shot protocol, DBT-Bleed outperforms all baselines. R2: S denotes similarity-score calculation; A is the spatial adapter from [17]. R3: Multiple IAEs and additional dataset testing are left for future work (Sec. 5). Clip-window N ablation is in Table 3.



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.

    DBT-Bleed proposes a dual-branch multi-scale temporal adapter for distinguishing active from residual bleeding, plus HiRED hierarchical frame selection, and introduces the EndoPit-IAE neurosurgical dataset. R3 accepts independently on clinical importance and zero-shot generalisation. R1 and R2 raise substantive concerns that the authors should address in rebuttal: (1) The central claim of distinguishing residual versus active bleeding is never analytically substantiated - there is no per-event analysis showing which events are correctly vs. incorrectly classified, nor is there a formal definition distinguishing the two classes. (2) Precision and AUPRC are not reported despite class imbalance. (3) The “any-frame-positive” labelling policy is more inflation-prone than majority vote. (4) The red-channel-based HiRED selection does not directly address the residual-vs-active distinction. (5) Annotator count, protocol, and inter-annotator agreement for EndoPit-IAE must be reported. The authors should address these issues thoroughly in the rebuttal.

  • 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 reviewing the rebuttal letter and reviewers’ post-rebuttal comments, I recommend acceptance as it majorly addresses the reviewers’ concerns.



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 authors have addresses key points raised by the 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.

    Accept

  • Please justify your recommendation.

    The reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns methodological applications in CAI.



back to top