List of Papers Browse by Subject Areas Author List
Abstract
Early screening via colonoscopy is critical for colon cancer prevention, yet developing robust AI systems for this domain is hindered by the lack of densely annotated, long-sequence video datasets. Existing datasets predominantly focus on single-class polyp detection and lack the rich spatial, temporal, and linguistic annotations required to evaluate modern Multimodal Large Language Models (MLLMs). To address this critical gap, we introduce Colon-Bench, generated via a novel multi-stage agentic workflow. Our pipeline seamlessly integrates temporal proposals, bounding-box tracking, AI-driven visual confirmation, and human-in-the-loop review to scalably annotate full-procedure videos. The resulting verified benchmark is unprecedented in scope, encompassing 528 videos, 14 distinct lesion categories (including polyps, ulcers, and bleeding), over 300,000 bounding boxes, 213,000 segmentation masks, and 133,000 words of clinical descriptions. We utilize Colon-Bench to rigorously evaluate state-of-the-art MLLMs across lesion classification, Open-Vocabulary Video Object Segmentation (OV-VOS), and video Visual Question Answering (VQA). The MLLM results demonstrate surprisingly high localization performance in medical domains compared to SAM-3. Finally, we analyze common VQA errors from MLLMs to introduce a novel “colon-skill” prompting strategy, improving zero-shot MLLM performance by up to 9.7% across most MLLMs. The dataset and the code are available at https://abdullahamdi.com/colon-bench .
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1043_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1043_supp.zip
Link to the Code Repository
https://github.com/ajhamdi/colon-bench-eval
Link to the Dataset(s)
https://huggingface.co/datasets/ajhamdi/colon-bench
BibTex
@InProceedings{HamAbd_ColonBench_MICCAI2026,
author = { Hamdi, Abdullah AND Yang, Changchun AND Gao, Xin},
title = { { Colon-Bench: An Agentic Workflow for Scalable Dense Lesion Annotation in Full-Procedure Colonoscopy Videos } },
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
+Agentic pipeline to create a new benchmark of 528 short video segments from 60 long videos from the REAL-COL dataset, totaling 464,035 frames,14 distinct lesion categories, over 300,000 bounding boxes, 213,000 segmentation masks, and 133,000 words of clinical descriptions. Each short video segment is associated with a lesion, segmentation masks, and bounding boxes. +The agentic pipeline includes 3 agents with human verification at the end. The three agents are (1) the video analysis agent that creates candidate video segments with a lesion and a corresponding text description; (2) the video verification agent to verify the work of the first agent; (3) the AI confirmation agent to verify the annotated bounding boxes and segmentation masks. +Evaluation of 13 MLLMs for lesion classification, Segmentation, and visual question answering (VQA); this includes three versions of Gemini (2.5, 3 Flash, and 3 Pro) +New prompt based on analysis of the most common errors by MLLMs understudied. +The new benchmark could help advance AI-assisted analysis and retrieval of colonoscopy videos.
- 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 offers the following.
- Clinical descriptions and additional lesion categories not available in the five datasets listed in Table 1
- Human verification of the short video segments annotated by the agentic pipeline. +Improved performance with the proposed “colon-skill” prompt by 9.7%. +A user interface for humans is provided.
- 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.Related State-of-the-Art: EndoBench accepted for publication in NeurIPS 2025 was missing. https://neurips. cc/virtual/2025/loc/san-diego/poster/121546 EndoBench includes colonoscopy videos with lesion-related questions. Please compare the proposed dataset qualitatively with EndoBench regarding colonoscopy. Open-source Medical Domain MLLMs were also evaluated in EndoBench. 2.Table 1: When listing the number of videos, it would be better to include the average and standard deviation of the video lengths in seconds. The 528 videos in Colon-Bench were derived from REAL-COL and are much shorter than the videos from REAL-Col. 3.Issues with clarity of methodology, involvement and experience of human annotators in verifying questions and answers, and clarity of explanation of the reported results. The details are below.
+Gemini-3 versions were used for creating candidate video segment proposals, video verification, and bounding boxes. What was the prompt used? What is the t_p value in the initial temporal proposal? +It is unclear how the recall in Table 2 should be interpreted. The recall after verification is 68.7.Does it mean that the agent misses about 21.3 of real lesion segments? Did the humans verify the correctness of the AI automated video verification step (the 2nd agent) to derive the recall and precision values? What kind of video segments did it miss? +What was the experience of the humans (e. g. , GI doctors, year of conducting colonoscopy, or the authors who have experience in this field), how many evaluators were used, and what were the interrater agreements among the evaluators if they rated overlapping video segments? Did they also verify the generated questions and answers? +Questions for VQA were generated using Gemini-3? Have the questions and answers been verified by humans? If so, did the expert agree with the questions and results? How difficult were these questions? The authors mentioned two difficulty tiers, but provided no further information. For instance, what proportion of these questions were about lesion sizes, abnormality types, or others? What prompt and template were used? A clinician-evaluator for a subset of QA can be conducted to verify the generated questions. See EndoBench.
+Gemini 3 wins in the VQA tasks. Could this be caused by bias in using Gemini-3 to generate these VQA pairs? Debiasing Colon-Bench was discussed, but the discussion did not give details on how adversarial distractors were generated. Were they generated by Gemini 3?
+Human-generated question-and-answer pairs should also be included to increase the value of the benchmark.
+Section 3.1: How many questions does the majority of the model answer incorrectly, and what kind of questions were they? This is to evaluate the “colon-skill” prompt. What does the prompt look like? Will this prompt generalized to a different dataset as well?
+No mention of whether this benchmark will be available for reproducibility tests.
Experimental setup:
- Evaluation on some recent open-source MLLMs, not provided. +Error analysis for patterns of errors not provided.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper has the potential to be highly impactful, there are several key weaknesses. The human verification step in the various steps is not clearly described. The question-answer construction description reads as if it does not involve human verification. The experience of the human evaluators was not described, but it is important for readers to know the quality of the ground truth for the medical domain. Reproducibility is another issue, as there is no mention of the availability of the dataset and the question-answer pairs or prompts used in the experiments. There is a potential bias toward Gemini-generated questions, with the best results coming from Gemini. A comparison with EndoBench, focusing solely on its colonoscopy subset, could be added in Table 1 to strengthen the dataset’s novelty and provide readers with a clear sense of the differences.
These issues can be addressed, but it will take time and significant revision to have domain experts verify a large enough subset of the 1,485 five-choice questions and the 2,740 generated questions.
- 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.
I’ve read the rebuttal, which answers several questions.
Positive points: (1) Eventually, the community will have 528 short clips from the 60 full colonoscopy videos from REAL-COL. These 528 short clips with annotations of 14 lesion types and clinical texts associated with these lesions, along with segmentation marks and bounding boxes. The community will have 4,225 multiple-choice questions/answers with 98% correct based on 10% sampling done by the authors. (2) The community will have the dataset, code, prompts, SKILL.md, and Jsons on Hugging Face and GitHub on acceptance.
Negative points:
- The ground truth is not as good as in EndoBench, which includes static images and questions from several colonoscopy datasets, compared with clinicians’ performance.
- The claim that the dataset has “representative full-procedure coverage” lacks evidence to support it since the clinician review of the detected short clips was done after AI filtering steps. We do not know whether any lesions were filtered out by the AI steps or not. Table 1 needs to be modified to avoid misrepresentation. Table 2 needs clarification about the numbers presented.
- Compared to EndoBench, the benchmark lacks the performance of open-source medical MLLM, such as ColonGPT.
- The criteria of difficulty of the questions are not indicated. The difficulty of questions play important role in MLLM performance evaluation.
Reasons for Rejection: The existing EndoBench on colonoscopy can be used to evaluate VQA performance of MLLMs. For segmentation and detection tasks, it is more important to test on full-length procedures rather than short clips. This is because there are fewer lesion segments than normal frames. Because there is no verification that the AI steps actually filtered out actual lesions, researchers cannot rely solely on this dataset as ground truth for segmentation or detection for the entire procedure.
Reasons for Acceptance: The dataset and code are promised to be available. There are 14 lesion types that were not in Real-Col. They can be useful for the community. The camera-ready version must ensure that all points raised are discussed and that their contributions are appropriately phrased (Table 1). This reviewer is also ok with it if the committee wants to accept it.
Review #2
- Please describe the contribution of the paper
This paper introduces Colon-Bench, a densely annotated colonoscopy video dataset via multi-stage agentic filtering from REAL-COLON. Various MLLMs are benchmarked on Colon-Bench for 14 lesion classification, open-vocabulary video segmentation, prompted and unprompted VQA. Additionally, the paper proposes a “Colon-Skill” prompt augmentation strategy to improve MLLM performance for colonoscopy VQA.
- 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.This paper aims at addressing the lack of densely annotated, video datasets beyond single-class polyp detection with free-text annotations, which is a critical bottleneck in current colonoscopy multimodal AI research. 2.The multi-stage agentic flow mostly leverages MLLMs and tracking models, effectively reducing manual efforts for dense video annotations. This pipeline design could benefit similar large-scale video dataset establishment. 3.The multi-class lesion and VQA benchmarks in Colon-bench are valuable contributions to multimodal colonoscopy studies. 4.The paper provides broad benchmarks across many SOTA MLLMs, providing useful insights for future MLLM applications in colonoscopy. 5.The proposed ‘Colon-skill’, which aggregates failure patterns across MLLMs, is an intuitive design for simple and robust prompt augmentation.
- 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.Despite the paper claiming a large-scale dataset, the proposed Colon-bench is derived directly from the existing REAL-COLON dataset via annotation refinement. This limits the scope of the dataset contribution. The final dataset comprises 528 clips from the original dataset, which does not support the claim of “full-procedure videos” in the introduction or conclusion. 2.The overly simplified multi-stage annotation pipeline in the method section lacks descriptions for key components to assess validity and reproducibility. For example, there is no description about how the “cued AI confirmation” processes videos, filters errors, or ensures annotation quality. 3.Some of the pipeline designs are based on existing work with domain-specific engineering, such as widely used error-driven prompting in ‘Colon-skill’, which should be properly cited in the paper. Such confusion makes it hard to assess the method’s novelty. 4.For the VQA, the paper does not clearly specify whether the free-text descriptions are fully generated by MLLMs, or the extent of clinical verification if experts are involved. 5.The reported dataset quality (precision: 55.4, recall:48.6, F1: 51.8) of the curated dataset is limited. Precision at 55.4% indicates that there are still many false positives after all the filtering steps. This level of noise raises concerns about benchmarking reliability. 6.In the MLLM benchmarking, the paper reports that Gemini models have the best performance. As the annotation pipeline relies heavily on Gemini, there may be bias not discussed in the paper.
- 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
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 proposed a multi-stage agentic pipeline to address the lack of densely annotated, multimodal video datasets beyond single-class polyp detection for colonoscopy research. The proposed Colon-Bench dataset, with multi-class lesion annotations, segmentation masks, free-text descriptions, and VQA tasks, is broadly useful for advancing video understanding and MLLM research. However, several important limitations require further clarification. The paper writing omits clear descriptions about key components of the annotation pipeline, raising concerns about reproducibility. More critically, the annotation quality (precision 55.4%, recall 48.6%) is relatively low for a reliable benchmark. The heavy reliance on Gemini models during annotation might introduce potential bias in the MLLM benchmarking results, which should be discussed in the paper. Despite these weaknesses, Colon-bench is still a useful dataset, and a benchmarking contribution is likely to benefit the community, placing the paper slightly above the acceptance threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 rebuttal doesn’t resolve the common concern of the evaluation bias towards Gemini models. And the paper needs a major revision to improve clarity of the method and experiments, including details on low precision/recall.
Review #3
- Please describe the contribution of the paper
- An annotated benchmark, named Colon-Bench, built by enriching the existing REAL-COLON dataset through an agentic annotation workflow. Colon-Bench extends the original data with a dense, multi-layered annotation schema covering multi-class lesion classification, bounding boxes, segmentation masks, natural language descriptions, and open-vocabulary video object segmentation
- A multi-stage agentic workflow, incorporating human-in-the-loop review, for comprehensive annotation of full colonoscopy procedures
- The first benchmark to address Open-Vocabulary Video Object Segmentation (OV-VOS) in colonoscopy, filling a notable gap in the evaluation landscape for this clinically relevant task
- A proposes two-stage prompt augmentation pipeline that injects structured domain knowledge into MLLMs at inference time, requiring no additional training, and demonstrating consistent accuracy improvements on colonoscopy understanding tasks.
- 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 introduction of a multi-stage agentic annotation pipeline that incorporates human-in-the-loop review to progressively filter and refine lesion detections. This is a practically valuable contribution, as producing reliable annotations for colonoscopy video at scale is a known bottleneck in the field.
- A stage-by-stage measurement of false positive reduction, which provides transparent and concrete evidence that each stage of the pipeline contributes meaningfully to annotation quality.
- Strong quantitative evaluation of the proposed Colon-Skill on several VLMs are clinically meaningful
- 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 described annotation pipeline is precision-oriented. It filters candidates through successive verification stages, including human review, to ensure the correctness of retained annotations. While this effectively reduces false positives, it does not address false negatives. Lesions never surfaced as candidates remain invisible to every subsequent stage, including human review, as no reviewer can confirm what they were never shown. This is a structural blind spot that the authors do not consider. Notably, the drop in recall in Table 2 is a direct consequence of this design choice, yet it is presented without discussion, and the way it is highlighted in the table is potentially misleading 2.Clarity issue concerns:
- The terms ‘videos’ and ‘windows’ are used interchangeably throughout the paper, which is misleading. The dataset is built upon the 60 videos of REAL-COLON, from which 528 curated windows are extracted. Presenting 528 as the dataset size and comparing it against REAL-COLON’s 60 videos in the comparison table creates a false impression of scale, as the two numbers refer to fundamentally different units
- Section 2.2 states that Colon-Bench comprises 1,597 clips, but this number is not clearly justified. Given that the curation pipeline yields 528 windows, it is unclear how this figure is derived and whether it is consistent with the rest of the paper.
- 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
Minor comments:
- typo in Subsection 2.1: “Colon-becnh”
- Table 3 - to my understanding, there are three benchmark tasks (not four): (1) VQA accuracy, (2) lesion classification and (3) video segmentation.
- 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 offers a useful annotation pipeline and the Colon-Bench benchmark, and its evaluation of Colon-Skill across multiple VLMs is clinically relevant. However, two issues reduce its impact. First, the pipeline favors precision and does not consider false negatives, which explains the recall drop in Table 2, but is not discussed and is presented in a potentially misleading way. Second, several numerical inconsistencies and unclear terminology throughout the paper, including in the comparison table, create confusion about the true scale and composition of the dataset
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors acknowledge the recall limitation and committed to clarifying it. While the false-negatives blind spot was not fully resolved, I find it acceptable as a known design tradeoff that can be disclosed in the paper. The concerns about terminology and metrics was addressed by the authors.
Author Feedback
We thank the reviewers; each section first quotes the criticism.
(1) “Gemini-3 wins, so the benchmark is biased toward Gemini” [R1, R2]. To refute the bias hypothesis, we tested a recently released non-Gemini control: open-source Gemma 4 31B achieves 63.8 / 59.9 / 85.1 / 47.2 on Prompted-VQA / Unprompted-VQA / Classification-Acc / Seg-mIoU, almost matching Gemini 3 Pro (top of Table 3) .
(2) “How are adversarial VQA distractors generated for debias?” [R1]. Distractors come from a Gemini-3 call seeing only the question and correct answer (not the video), emitting 4 length, style, and specificity-matched alternatives; positions are re-randomized. Blind text-only LLM accuracy drops from ~70% (naive distractors) to ~20% (random), confirming text cues are not exploitable.
(3) “Human-expert verification and VQA content are under-specified” [R1, R2]. All windows were reviewed by the authors with Dr. J.H., a surgeon with over 30 years of GI/colonoscopy practice (11.6% expert-rejection, Fig.1). Lesion descriptions are AI-generated by the verification agent; the expert accepts/rejects the window+annotations+description package. Two tiers: Prompted (bbox, no time refs) and Unprompted (no bbox, time-referenced queries); questions cover lesion type, size, location, interventions, and process descriptions. Up to 10% of VQA pairs were manually checked by the authors with 98% correctness (supp. mp4).
(4) “Table 2 Precision/recall 55.4 / 48.6 indicate a noisy benchmark” [R1, R2, R3]. Specificity 94.4% is the relevant reliability indicator. The 55.4 / 48.6 figures compare our 14-category annotations against REAL-COLON’s polyp-only labels from a commercial detector flagging ~12% of REAL-COLON’s frames; most of our categories (ulcers, bleeding, angiectasia, etc.) are absent from this reference, so the labels mismatch by definition. Table 2 will be reframed for R3’s false-negative concern; we target representative lesion-spectrum coverage, not exhaustive frame-level recall.
(5) “No release commitment” [R1, R2]. Dataset, code, prompts, SKILL.md, and Jsons will be released on Hugging Face and GitHub on acceptance. Detailed methodology was in a supplementary PDF disallowed by MICCAI 2026 this year and will fold into the camera-ready.
(6) “Missing EndoBench (NeurIPS 2025)” [R1]. EndoBench (Dec 2025) is concurrent with our Feb 2026 submission and will be cited. It evaluates static-image VQA on general endoscopy. Colon-Bench is the first video colonoscopy benchmark: dense video segmentation, binary lesion classification (over a 14-category lesion spectrum), and prompted/unprompted video VQA. Disjoint scope.
(7) “Inconsistent units: 60 vs 528 vs 1,597 in Table 1 & Sec. 2.2 “ [R1, R2, R3]. 60 source recordings are full-procedure videos scanned end-to-end; 528 = annotated lesion windows extracted from them (duration: median 59.9 s, mean 87.9 s, std 83.2 s); 264 clips in segmentation benchmark. 1,597 = VQA clip pool: Prompted (bbox) + Unprompted (no bbox) renderings of accepted windows + confirmed non-lesion distractors. Table 1 will add duration stats; Sec.2.2 will spell out the 528/264/1,597 relation; “full-procedure” softened to “representative full-procedure coverage”.
(8) “Pipeline and Colon-Skill are under-specified” [R1, R2]. Each agent uses a JSON-templated prompt; t_p = window duration (s) for temporal grounding. The AI-cued confirmation agent ingests the clip with bbox+description overlay and outputs accept/reject + rationale. Colon-Skill aggregates up to 20 common errors per lesion category (a design choice) into anti-error rules (e.g., “a flat lesion on a haustral fold without a fibrous tether is Sessile”). +9.7% gain across model families.
In sum: bias is refuted (Gemma 4 control + debiasing Q/A); verification was by a 30+ year senior surgeon, including the VQA check; reproducibility committed; precision/recall are cross-taxonomy. We respectfully ask R1 to reconsider, citing their own “potential to be highly impactful” assessment.
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 reviewers commonly raise concerns that the dataset contribution and scale are overstated, since Colon-Bench is derived from REAL-COLON and appears to consist of curated clips/windows rather than full-procedure videos, with unclear and inconsistent reporting of dataset units and counts. They also emphasize that the annotation and VQA generation pipeline lacks sufficient methodological detail, human-expert verification, reproducibility information, and error analysis, while the relatively low reported annotation quality and heavy reliance on Gemini may introduce noise and bias into the benchmark.
- 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.
Most of the concerns raised by the reviewers (human expert confirmation, Bias on Gemini, …) were addressed in the author’s feedback.
Dataset, code, prompts, SKILL.md, and Jsons will be released on Hugging Face and GitHub on acceptance. Furthermore, the authors will make their dataset and codes publicly available. It will contribute to the community’s progress.
This manuscript is worthy of acceptance.
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.
This paper received Reject, Reject, and Accept.
The paper makes a real contribution in dataset curation and annotation: 528 lesion windows extracted from REAL-COLON’s 60 full-procedure videos, spanning 14 lesion categories (vs prior single-polyp datasets) with 300k bounding boxes, 213k segmentation masks, and 133k words of clinical descriptions, and the dataset is publicly released; however, the two post-rebuttal Rejects raised structural concerns that the rebuttal did not fully resolve: the pipeline-evaluation circularity (Gemini models generate the lesion proposals, bounding boxes, lesion descriptions, and VQA questions, and Gemini 3 Pro/Flash then top all four benchmark tasks), the dataset-size misrepresentation in Table 1 (528 lesion windows compared against REAL-COLON’s 60 full-procedure videos), and the limited human verification of the AI-generated VQA pairs (10% sample by the authors themselves).
My decision to accept rests on the dataset’s tangible value to the community (14 lesion categories vs prior single-polyp datasets, dense multi-modal annotations, and confirmed public availability), and is conditional on the authors severely addressing the remaining concerns in the camera-ready.
The authors must (1) explicitly disclose the Gemini-dependence throughout the annotation and VQA-generation pipeline and discuss the resulting benchmark bias as a known limitation; (2) correct Table 1 to compare like with like (lesion windows vs lesion windows, full-procedure videos vs full-procedure videos) and add duration statistics for the 528 windows; (3) acknowledge the limited human verification (10% sample by the authors) as a known limitation rather than a quality endorsement; (4) add EndoBench (NeurIPS 2025) to the related work and discuss the disjoint scope (static images vs video); and (5) reframe the 55.4/48.6 precision/recall in Table 2 to make the cross-taxonomy mismatch with REAL-COLON’s polyp-only labels explicit.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The reviewer points out that EndoBench is suitable for VQA but inadequate for segmentation and detection due to the need for full-length procedure testing and unverified AI filtering, while also noting that the rebuttal failed to address evaluation bias towards Gemini models and that the paper requires major revisions to improve methodological clarity and explain low precision/recall results.