List of Papers Browse by Subject Areas Author List
Abstract
Temporal vision-language models (TVLMs) offer a reusable, prompt-based interface for surgical video understanding, yet, their robustness under clinically realistic acquisition artifacts in endoscopy remains insufficiently characterized. In practice, degradations such as defocus, haze, motion blur, noise, cautery smoke, and packet loss introduce structured distribution shifts which may compromise video–text alignment. We study the robustness of temporal VLMs under such shifts caused by corruptions in clip frames.
We introduce Endo-C6, a compact corruption benchmark of six endoscopy-realistic perturbations evaluated at a fixed high severity, and apply it to public Gastrointestinal (GI) endoscopy and laparoscopic cholecystectomy videos.
Under a standardized prompt protocol, we benchmark 3 recent surgical TVLM baselines and analyze robustness in both mean and worst-case settings, spanning 294 dataset-level evaluations. Finally, we present RobustEndoCLIP, obtained by few-shot parameter-efficient tuning with VeRA,
outperforming existing TVLM baselines.
Our findings show that off-the-shelf TVLMs can exhibit severe worst-case collapse under endoscopy-specific corruptions, whereas lightweight few-shot adaptation can substantially improve corrupted performance and robustness without changing the prompt-based interface. We expect Endo-C6 to support standardized robustness reporting and promote more reliable clinical vision-language systems.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4661_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DarakshanRashid/RobustEndoCLIP
Link to the Dataset(s)
https://github.com/DarakshanRashid/RobustEndoCLIP
BibTex
@InProceedings{RasDar_On_MICCAI2026,
author = { Rashid, Darakshan AND Imam, Raza AND Khan, Ufaq AND Bilal, Muhammad AND Ashraf, Shazad AND Mahapatra, Dwarikanath AND Yaqub, Mohammad AND Khan, Muhammad Haris AND Razzak, Imran AND Lall, Brejesh AND Maier-Hein, Lena AND Xie, Yutong},
title = { { On the Robustness of Temporal Vision-Language Models for Surgical Endoscopy Videos } },
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
This paper presents a robustness benchmakr for endoscopy named Endo-C6, which consists of six clinically realistic endoscopy corruptions and desgined to model acquisition artifacts that occur in real GI endoscopy and laparoscopic surgery. It also propose a standarlized prompt-based evaluation protocol for testing surgical TVLMs under these controled shifts across public datasets. Finally, it propose RobustEndoCLIP, a lightweight few shot VeRA-based adaptation of CLIP-style temporal model that keeps the same prompt interface while substantially improving performance on corrupted data.
- 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 motivation on surgical endoscopy is noisy and investigate the robustness motivation is clear and sound, especially the paper emphasizes worst case failure rather than only average performance. 2.The paper benchmarks three recent surgical TVLMs across three datasets, seven conditions (clean + six corruptions), and two corruption-coverage settings, for 294 dataset-level evaluations and 593,488 clip-level evaluations. These contribute to a large-scale study.
- 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.Although I appreciate the benchmark setting, the methodology novelty is limited if author presented this paper as method paper. The underlying common-corruption benchmark paradigm was established by ImageNet-C (Hendrycks & Dietterich, 2019), which already evaluates multiple corruption types across five severity levels. In medical VLMs, Imam et al. (2025) already introduced MediMeta-C and RobustMedCLIP, combining a corruption benchmark with few-shot adaptation for robustness. In endoscopy specifically, Wang et al. (2024) already benchmarked robustness under synthetic corruptions for depth estimation, and VeRA itself is an existing PEFT method from Kopiczko et al. (2024). So the real novelty here is narrower: applying these ideas to temporal surgical TVLMs with an endoscopy-specific corruption set and analysis. That is still useful, but it is not a fundamentally new robustness formulation. 2.The introduction motivates prompt-based, open-vocabulary recognition, but RobustEndoCLIP is trained with cross-entropy on a fixed in-dataset label set and is evaluated only on those same seen vocabularies. So what is shown is stronger closed-set robustness with text prompts, not robustness of unseen-label transfer or open-vocabulary generalization after adaptation. A held-out-label, cross-dataset prompt-transfer, or retrieval evaluation after tuning would be needed to support that stronger claim.
- 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?
I think the paper would be position much better as an application paper rather than a methodology paper.
- 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.
After reading author’s rebuttal, I am happy to raise to accept.
Review #2
- Please describe the contribution of the paper
This paper introduces Endo-C6, a benchmark for robustness evaluation of endoscopic TVLMs under realistic corruptions; it systematically benchmarks three recent surgical TVLMs on this setting; and it proposes RobustEndoCLIP, a few-shot parameter-efficient adaptation method that improves robustness under corrupted conditions.
- 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 addresses a clinically meaningful and underexplored problem: the robustness of surgical TVLMs under realistic endoscopic corruptions. It introduces a well-motivated corruption benchmark spanning six clinically relevant perturbations. The proposed RobustEndoCLIP is practically interesting because it shows that lightweight few-shot parameter-efficient tuning can substantially improve robustness without changing the prompt-based deployment interface.
- 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 technical novelty of the proposed method appears limited. RobustEndoCLIP is essentially built on top of an existing temporal CLIP-style model and adapts it using VeRA, which is itself an existing parameter-efficient tuning method rather than a new methodological contribution of this paper. 2.The baseline comparison is somewhat narrow. Although the paper evaluates three recent TVLMs, these baselines all come from a closely related line of surgical video-language pretraining work, rather than representing a broader spectrum of temporal vision-language models. As a result, it is still unclear whether the observed conclusions would generalize to a wider set of architectures, including stronger general-domain video-language models or additional medical/surgical multimodal models. 3.The evaluation is still limited in scope, being centered on prompt-based classification with a fixed corruption setting, which makes the robustness claims less comprehensive than they could be. 4.Additional weaknesses are that the “few-shot” setting still uses 16% labeled training data, the corruptions are synthetic rather than real clinical degradation, and the paper provides limited analysis of why VeRA improves robustness beyond reporting the gains.
- 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?
The main contribution is a clinically motivated robustness benchmark and analysis, while the proposed method is technically incremental.
- 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
Review #3
- Please describe the contribution of the paper
In this work, the authors attempt to answer four scientific questions about the robustness of endoscopic visual language models. A Corruption Benchmark–Endo-C6 was defined, and 3 temporal vision-language models (TVLM) were tested. In addition, a few-shot adaptation method, RobustEndoCLIP, was presented to improve the performance of TVLM under various corruptions.
- 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.
This study sought to assess the robustness of Temporal Vision-Language Models in the surgical endoscopic setting. The authors compiled three open-source endoscope datasets and conducted a taxonomic study on different types of endoscopic video corruptions. This work conducted empirical studies on three start-of-the-art TVLM models and two PEFT fine-tuning methods.
- 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.
My main concern is the lack of methodological innovation. The corruption taxonomy is more like common sense rather than the author’s insight/contribution. The dataset is from other open-source works without any re-annotation. Meanwhile, I am not confident that the benchmark experiment is comprehensive or solid. There is no innovation regarding parameter-efficient fine-tuning. In addition, this paper does not specify what task these TVLMs are intended for.
- 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?
This work is more like an empirical study/benchmark of the existing TVLM model for endoscopic videos. I think this type of work needs to meet higher standards to be accepted by MICCAI. I suggest Weak Rejection.
- Reviewer confidence
Somewhat confident (2)
- [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 author’s feedback emphasizes clinical motivation and clarifies details. However, I am not convinced that this benchmark is comprehensive enough to be accepted.
1) Dataset: There are many endoscopic datasets. As a benchmark study, experiments on only 3 datasets are not sufficient. In https://arxiv.org/pdf/2603.27460, the survey includes 76 endoscopic datasets. I recommend adding more datasets/tasks/methods and submitting to a journal. 2) Taxonomy: I still do not think this can be counted as the author’s main contribution… Everyone working in surgical computer vision is quite familiar with these corruptions and mentions these in almost every paper… 3) Baselines: Similar to datasets, please consider including more models to give a comprehensive benchmark. I personally do not accept a benchmark paper that only tests 3 baseline models. 4) Concept: What is the definition of VLM robustness? This study does not provide concrete quantitative metrics to reflect this term. At the beginning, there are 4 questions, but it is really hard to follow them while reading the last parts of the paper. The presentation of this paper can be improved from my perspective.
Author Feedback
We appreciate the reviewers’ effort and constructive comments. We are glad that R1 recognized the clear motivation, worst-case focus, and large-scale evaluation; R2 highlighted the clinically meaningful Endo-C6 setting and practical robustness gains; and R3 noted the empirical study of TVLM robustness, taxonomy, datasets, and TVLM/PEFT evaluation. Below, we address the concerns point-by-point and will incorporate clarifications in the revision.
R1.1/R2.1 Contribution scope relative to prior work. We appreciate R1/R2’s framing and will emphasize the paper’s central contribution: a clinically driven Endo-C6 robustness study for temporal surgical TVLMs, with RobustEndoCLIP as the limited-supervision adaptation setting. Prior work addresses natural corruptions, medical VLM robustness, or endoscopic depth estimation, but not prompt-based surgical TVLMs under GI/laparoscopic acquision artifacts. To our knowledge, Endo-C6 is the first benchmark for this setting, combining paired label-preserving corruptions on real clips, domain/temporal artifact structure, limited-supervision analysis, and mean/worst-case reporting. It reveals artifact-specific failures: smoke dominates laparoscopy, GI sensitivity is model-dependent, and clean accuracy poorly predicts corrupted robustness.
R1.2 Prompt-based vs. open-vocabulary scope. We agree experiments use fixed dataset vocabularies, not held-out-label transfer. Here, “prompt-based” means CLIP-style class-name prompts matched with video embeddings. Claims concern label-preserving corruptions under this interface, not unseen-label adaptation after tuning. We will clarify the task as CholecT50 phase, Kvasir GI landmark/pathology, and TEMSET-24K action classification.
R2.2 Baseline selection. SurgVLP, HecVLP, and PeskaVLP are recent public surgical TVLMs with comparable CLIP-style interfaces and no Endo-C6 exposure. Identical prompts, sampling, splits, corruptions, and temporal coverage enable controlled comparison; broader VLMs/surgical multimodal models are valuable future works.
R2.3/R3.3 Task scope and benchmark strength. Endo-C6 evaluates prompt-based temporal clip classification: CholecT50 phases, Kvasir GI landmark/pathology classes, and TEMSET-24K actions. Within this scope, it is comprehensive and solid: full public GI/laparoscopic test splits, 3 datasets, 3untuned TVLM baselines plus RobustEndoCLIP configurations, clean+6 corruptions, 2 temporal coverages, and 294 dataset-level/593,488 clip-level evaluations under identical prompts, sampling, splits, severity, and aggregation. This controlled design isolates video-text alignment failures under label-preserving shifts and avoids broader claims. We will further clarify the task/scope and make it explicit in the final, and extend the scope in future work.
R2.4 Few-shot, corruptions, and VeRA. The 16% setting is the largest limited clean-label budget; Fig. 4 reports 0/4/8/16%. Synthetic corruptions are deliberate: applied to real test clips, they keep labels/test composition fixed while varying only artifact type/severity, simulating routine endoscopic acquisition artifacts for reproducible paired stress tests. Table 2 shows LoRA can improve clean accuracy, but VeRA gives stronger corrupted performance with far fewer params (1,024 vs. 720,896; 29.13 vs. 16.92 average Mean-C), suggesting better regularization under corruption.
R3.1 Methodological scope and taxonomy. The core contribution is a clinically motivated robustness benchmark and analysis for temporal surgical VLMs, not a new adaptation algorithm. Endo-C6 is not a generic corruption list: organizing clinically familiar artifacts by domain salience and temporal behavior enables interpretable smoke/laparoscopy and motion/defocus/GI findings.
R3.2 Public datasets and re-annotation. We use public datasets to ensure reproducibility and isolate artifact-driven robustness failures. Since Endo-C6 applies label-preserving corruptions to existing test clips, re-annotation is not required.
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.
Borderline case, where the biggest concern is the lack of novelty. However, the reviewers do mention the merit of an application-based benchmark style paper such as this one, hence I would suggest a rebuttal to break ties.
- 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.
All three reviewers acknowledged a strong rebuttal, one reviewer indicating that she/he is even more confident about proposing to accept the paper. Some issues remain, as pointed out by reviewer 3 (who propose to reject): Number of data sets used, number of baselines, and limited scope of corruptions (R2). However, there’s only so much you can fit into a MICCAI paper and suggest the authors still extend the work and submit to a journal, as reviewer 3 suggests, but also present the current results at MICCAI, also supported by 2 of 3 reviewers propose an accept, compared to a Weak Reject by the other reviewer.
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 idea presented in the paper is interesting, and robustness evaluation is critical in our field. The rebuttal addressed the major 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 answers provided by the authors in the rebuttal letter are complete.
