List of Papers Browse by Subject Areas Author List
Abstract
Vision–language models (VLMs) show strong potential for automating endoscopy‑video interpretation, including detailed description, disease‑severity assessment, and temporal event localization. However, most surgery and endoscopy focused VLMs remain trained primarily on static images, despite reliable assessment requiring long‑range temporal reasoning over continuous procedures. This modality mismatch limits their applicability in real clinical workflows. Existing surgical video VLM datasets rely on noisy, public sources with limited clinical reliability or contain only short clips that fail to capture full procedural context. To address these gaps, we curated a long‑video dataset of endoscopy recordings from a clinical trial of patients with Inflammatory Bowel Disease (IBD). We additionally introduce AdaBIMBA, a novel adaptive token compression module comprising (i) a Token Scorer that predicts frame importance and (ii) a Mamba‑based compressor that adjusts per‑frame compression ratios accordingly. AdaBIMBA enables processing substantially more frames under the same memory budget while preserving clinically relevant content, achieving state‑of‑the‑art performance across all question categories on our IBD dataset.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2765_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YunKa_AdaBIMBA_MICCAI2026,
author = { Yung, Ka-Wai AND Mobadersany, Pooya AND Parmar, Chaitanya AND Chaitanya, Krishna AND Gunderson, Fabio AND Surace, Lindsey AND Ghanem, Louis R. AND Mansi, Tommaso AND Cula, Gabriela Oana AND Standish, Kristopher AND Damasceno, Pablo F.},
title = { { AdaBIMBA: Adaptive Token Compression for Long Endoscopy Video Understanding } },
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 paper introduces AdaBIMBA, a method designed to reduce the number of query tokens required to analyze a full-length colonoscopy video and answer 6 question types (2 grounding question types, 2 summarization question types, disease evaluation scoring, and reasoning). AdaBIMBA improves upon BIMBA by allocating more query tokens to video segments that contain information relevant to the question and fewer tokens to segments that do not. A query token captures important features of prior video frames rather than retaining original tokens per frame. AdaBIMBA was trained on 441 training videos and 10,153 QA-pairs. The test dataset has 140 videos with 3,545 question-answer pairs. The question-answer pairs were generated by GPT-4o.
- 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.
- AdaBIMBA outperforms 7 other models on this test set over different question types. These models do not include GPT-4o that was used to generate the questions.
- The 581 full-length video dataset for training and testing was annotated by domain experts, with boundaries for six anatomical segments and the Mayo Endoscopic score for each segment.
- The technical novelty is in the token scorer, which assigns a high score to important frames and a low score to less important frames. More query tokens are allocated to the high-score frames, and vice versa.
- An ablation study was done to show the impact of each component of AdaBIMBA.
- The presentation is nicely done, with diagrams showing the components of AdaBIMBA, example questions and answers, question types, video length distribution, and the distribution of different question types.
- 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 paper does not list the domain experts’ experience (e. g. , years of experience or the number of colonoscopies performed) who did the labeling. Did multiple domain experts review the same video segments? This information would help readers judge the quality of the dataset.
-
There was no mention of whether the LLM-generated questions were verified by the domain experts. A sample subset of the question set can be verified by domain experts.
-
There was no statistically significant testing to show that the improvement of AdaBIMBA is statistically significant.
-
It would be good to have the distribution of the number of anatomical segments with different MES scores of 0, 1, 2, and 3.This would help with the evaluation of Table 3.
-
Because the number of questions for each category is not the same, how is the average computed? Does each category get the same weight?
-
There was no measurement of how many tokens were actually reduced when compared to BIMBA.
-
The paper does not mention whether the code and dataset will be available, which will enhance reproducibility and impact of the work.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not 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
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?
The paper introduces a novel methodology and shows that it performs well compared to the existing methods under study. The proposed method outperformed the compared methods against one dataset, their own dataset. The presentation includes nicely done diagrams. The training and test datasets were created by the authors.
The paper could benefit from additional clarifying sentences to address the above drawbacks, such as listing the domain experts’ experience (e.g., years of experience or the number of colonoscopies performed), and interrater agreement among multiple raters when appropriate. The question-answer pairs were auto-generated, and there is no indication of whether they were verified by domain experts. Expert evaluation of question-answer pairs, even in a smaller subset, is common in prior benchmarking work. Sharing the question-and-answer pairs and the code would also improve reproducibility, even if the dataset cannot be shared.
- 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 satisfactorily answered the questions raised. Statistical tests were done. Dataset distribution for different MES scores is given.
Review #2
- Please describe the contribution of the paper
The main contribution of this paper is twofold. First, it introduces a long-form endoscopy video benchmark derived from an IBD clinical trial, with full-procedure recordings, segment-level annotations, severity labels, and multiple QA-style tasks. Second, it proposes AdaBIMBA, which extends BIMBA’s fixed compression scheme with an adaptive mechanism that first estimates frame importance and then allocates compression accordingly, allowing the model to process more frames under the same memory budget and achieve stronger overall results on the proposed benchmark.
- 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 has a number of appealing aspects. The problem setting is important and clinically relevant, as the work moves beyond static images or short clips and instead considers full-procedure endoscopy videos, which is closer to how these data are encountered in practice. The proposed method is also reasonably motivated. The authors build on the observation that clinically important abnormalities in endoscopy are often brief and sparse, and use this to justify an adaptive compression strategy based on frame importance; this is a sensible design choice for the task. I also appreciate that the paper does not rely solely on a single end-to-end result: the two-stage training procedure, the question-conditioned scorer, and the ablations suggest a genuine attempt to examine the contribution of different components. In addition, the evaluation covers several task types, including summarization, severity scoring, reasoning, and temporal grounding, which makes the empirical study fairly broad.
- 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 Introduction does not provide a sufficiently complete discussion of the most relevant literature on long-video compression and adaptive selection. Characterizing prior approaches mainly as “uniform frame sampling” is somewhat inaccurate and unfair, as many existing works already explore adaptive sampling or adaptive compression. I would encourage the authors to present a more balanced review and clarify the actual distinction of this work from prior studies. 2.The idea of first estimating importance and then allocating compression accordingly is intuitive and well motivated. However, the current comparisons do not convincingly isolate the contribution of the proposed method itself. AdaBIMBA uses a maximum of 1024 frames, compared with 512 for BIMBA and 128 for Qwen3-VL. While the average score improves from 50.8 to 54.7 over BIMBA, this gain is obtained with twice as many visible input frames. Although the paper repeatedly emphasizes the “same memory budget,” it does not yet provide sufficient evidence that the improvement mainly comes from the adaptive allocation strategy rather than simply from seeing more frames. 3.Another concerning observation is that AdaBIMBA performs worse than BIMBA on the most severe category, with MES=3 dropping from 53.5 to 49.6.Since higher MES corresponds to more active inflammation, and the paper positions clinically realistic endoscopy understanding as a key motivation, this degradation on the most critical cases deserves closer attention. I would encourage the authors to provide a more careful analysis of why the method underperforms on severe cases and discuss the potential implications.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I assigned a Weak Reject because I find the paper promising, but the current evidence does not fully support its methodological claims. The main issue is that the experiments do not cleanly disentangle the benefit of the adaptive compression strategy itself from the benefit of simply seeing more frames after compression: AdaBIMBA uses up to 1024 frames, whereas BIMBA uses 512, so the current results do not yet convincingly show that the gain mainly comes from adaptive allocation rather than increased visual coverage. In addition, the related-work discussion is somewhat incomplete, and the methodological advance over BIMBA feels more like a targeted extension than a clearly strong standalone novelty claim. Finally, the drop on the most severe category (MES=3) also deserves a more careful analysis. For these reasons, I see the paper as slightly below the acceptance bar at this stage, though I would be open to being persuaded by a strong rebuttal with better-controlled comparisons, a clearer positioning of the contribution, and a more convincing explanation of the observed failure mode.
- 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 construct a long video dataset of endoscopic recordings from clinical trials involving patients with inflammatory bowel disease (IBD). They further propose AdaBIMBA, which incorporates an adaptive token compression module consisting of (1) a token scorer that estimates the importance of each frame and (2) a Mamba-based compressor that adjusts the compression ratio accordingly. The proposed method enables processing of significantly more frames under the same memory budget while preserving clinically relevant information. It achieves state-of-the-art performance across all query categories on the IBD 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.
1.The main strength of this study is the construction of a long-duration endoscopic video dataset derived from clinical trials. Unlike conventional short, fragmented data, it captures the full examination process with high-quality expert annotations, enabling learning of full procedural context. 2.Another key strength is that the proposed method performs adaptive, frame-wise compression based on estimated frame importance, rather than relying on uniform frame subsampling. This enables efficient processing of long sequences while minimizing information loss.
- 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 description of the dataset is insufficient. In particular, the class distribution of the Mayo score, the data split proportions, and whether the split was performed at the patient level are unclear. These factors directly affect the reliability and reproducibility of the evaluation and should be clarified. 2.It is unclear to what extent the estimated importance aligns with clinically relevant content. The relationship between the model’s internal importance scores and clinical utility is not well established. Furthermore, the potential risk that adaptive frame compression may lead to the omission of clinically relevant information is not sufficiently discussed. 3.The efficiency gains from the proposed frame compression are not clearly quantified. In particular, reductions in computational cost, such as memory usage and inference time, are not sufficiently demonstrated. 4.While AdaBIMBA shows overall performance improvements, the reason for the performance decline in MES 3 is not explained. The analysis is insufficient, particularly regarding potential factors such as class imbalance and its interaction with frame compression.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not 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
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 adaptive compression based on importance scores for long-duration endoscopic videos, demonstrating novelty and practical relevance. However, it lacks sufficient validation of the relationship between importance and clinical utility. In addition, the descriptions are not sufficient to ensure reproducibility or to clearly quantify computational efficiency.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal adequately addressed most major concerns, particularly regarding dataset construction, patient-level splitting, class distribution, computational efficiency, and the additional MES=3 subgroup analysis. The authors also provided quantitative evidence for inference speed improvements and clarified reproducibility details.
The remaining concern is the limited validation of whether token importance truly aligns with clinically meaningful findings and whether adaptive compression may omit relevant information. While this limitation is acknowledged, additional expert-based clinical validation would further strengthen the claim.
Overall, the rebuttal substantially strengthened the paper and supports a more positive evaluation.
Author Feedback
We thank the AC and reviewers for their constructive feedback. We appreciate recognition of our method’s novelty(R1,3), clinical relevance(R2,3), broad evaluation and ablations(R1,2).
1) R2—Are gains simply due to more frames? To isolate frame-count from adaptive compression, we trained/evaluated AdaBIMBA with the same 512-frame input as BIMBA. AdaBIMBA scores 66.6/67.4/55.4/40.5/37.3/47.8 on Local/Global/Scoring/Reasoning/Event/Segment, outperforms BIMBA across all categories and improvs average 50.8 → 52.5 (paired t-test p=0.0045), confirming gains are not solely due to higher frame count. At 1024 frames, AdaBIMBA shows its second advantage: expanded temporal coverage under the same 2×80GB memory budget. BIMBA compresses 512 frame-tokens to ~128; AdaBIMBA compresses 1024 to ~132, maintaining nearly identical downstream token counts while doubling temporal coverage, further improving average to 54.7.Gains arise from both adaptive allocation at matched frame count and broader procedural coverage from efficient token utilization, without additional memory.
2) R1,2,3—MES=3 subgroup analysis The MES=3 drop is not statistically significant (p=0.335) with overlapping CIs. In different seeds AdaBIMBA (52.9,54.7) also matches or exceeds BIMBA (51.8,50.3), similarly under matched 512-frame (54.7 vs 53.5), suggesting variance in a smaller subgroup rather than systematic degradation. In contrast, AdaBIMBA shows significant improvements for MES=0–2 (p=0.001/0.010/0.005) and overall F1 (p<0.005), indicating stronger broader-cohort performance. A plausible explanation is that mild–moderate disease has sparse/localized findings benefiting more from adaptive allocation, while severe disease often involves diffuse pathology already captured in shorter windows. We will clarify this limitation and add future validation on larger MES=3 cohorts.
3) R2—Positioning relative to BIMBA and prior adaptive methods We revised Related Work to clarify AdaBIMBA’s contributions: (i) learnable question-conditioned Token Scorer for fine-grained per-frame adaptive compression; (ii) gradient-supervised two-stage training compatible with FlashAttention. Unlike heuristic saliency methods (e.g. LongVU), AdaBIMBA learns task-specific importance via QA supervision; unlike global compression (e.g. Dynamic-VLM), it adapts compression per frame; unlike keyframe selection (e.g. AKS), it preserves frames while reallocating token budget toward clinically relevant regions.
4) R1,3—Statistical significance, average weighting, efficiency Paired t-tests show significance across all categories (Event p<0.02; others p<0.005). Averages are unweighted, with consistent gains in weighted average (46.8→50.3). Token Scorer adds only ~1.21% parameters. Under identical pre-compression frame input generating 128 tokens, Qwen3-VL/BIMBA/AdaBIMBA take 28.7/12.7/6.54s (77% and 48.5% speedup; see 1) for token efficiency).
5) R1,3—Dataset details/reproducibility The dataset is split at patient level into 10,153/3,545 QAs from 441/140 train/test videos, preventing leakage. MES 0-3 counts are: train 766/562/539/467, test 280/185/212/166.Each video is annotated by up to 6 clinicians (3–8+ years experience, 59.7% MES inter-rater agreement). Scoring/Reasoning/Segment QA labels are directly derived from expert annotations; clinician validation of other QA categories is ongoing. Due to regulations, full data access will be case-dependent; but we will release detailed pseudocode, QA samples, dataset-construction and implementation details for reproducibility.
6) R3—Token importance and omission risk Fig 5 is intended to show the reliability of AdaBIMBA’s adaptive allocation rather than direct clinical utility. It shows less compression in question-relevant video regions, grounding answers in correct temporal evidence. While any compression may omit information, AdaBIMBA mitigates this vs. uniform BIMBA by retaining more information in important frames, instead of treating all frames equally.
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 comparison with BIMBA is confounded by frame budget (1024 vs. 512), so it is unclear whether gains come from adaptive allocation or simply from more visible frames. Second, the drop on MES=3 needs a careful explanation, including any interaction with class imbalance. Third, several dataset clarifications are needed: MES class distribution, train/test split proportions, and whether splits are patient-level, annotator experience and inter-rater agreement, and whether the generated QA pairs were verified by clinicians. Finally, the positioning relative to prior work on adaptive sampling and compression should be clarified.
- 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 provided a strong rebuttal and answered the previously raised points. The paper is well-written and the method thoroughly validated.
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 initial reviews nevertheless identified several substantive weaknesses. R2 raised the most important methodological concern: the main comparison with BIMBA was confounded by different visible-frame budgets, since AdaBIMBA used up to 1024 frames whereas BIMBA used 512.This made it difficult to determine whether the gains came from adaptive allocation or simply from seeing more frames. R2 and R3 also noted the degradation on the MES=3 subgroup, which is clinically important. R1 and R3 requested additional dataset information, including patient-level splitting, MES class distribution, annotator experience, inter-rater agreement, and whether LLM-generated QA pairs were clinically verified. Additional concerns included limited positioning with respect to prior adaptive compression/sampling methods, lack of statistical testing, limited efficiency quantification, and incomplete reproducibility details. The rebuttal substantially improved the assessment. The authors clarified that AdaBIMBA also outperforms BIMBA at the same 512-frame input budget, reported statistical testing, explained the 1024-frame setting as increased temporal coverage under a similar downstream token/memory budget, and provided additional analysis of the MES=3 subgroup. The rebuttal also clarified the patient-level split, MES distributions, annotator experience, inter-rater agreement, average weighting, inference time, and release plan for pseudocode/QA samples. These clarifications directly address many of the concerns raised by R1, R2, R3, and the initial meta-review. However, some of these clarifications include additional analyses not fully present in the submitted paper, so they should be interpreted primarily as evidence that the original claims are plausible rather than as a complete replacement for clearer controls in the manuscript. The remaining limitations are important but not sufficient to warrant rejection.
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 proposes a novel method, and its performance was demonstrated through the experiments. Therefore, I believe this paper is worthy of acceptance.